---
kind: concept
slug: imitation-learning
title: Imitation learning
summary: >-
  Learning a policy from recorded behaviour instead of a hand-written reward: behaviour cloning, nearest-neighbour lookup, return-conditioned sequence models, imitation terms in offline RL, vision-language-action and tactile policies, planning from logged video, imitating solver branching, expert hints for LLM agents and style imitation, with the trade-offs in data, compute and evaluation.
tags: [robot-learning, behavior-cloning, self-driving, vision-language-action, offline-rl, tactile-sensing, combinatorial-optimization, human-in-the-loop, legged-locomotion, navigation]
papers: ["1810.02890", "1906.01629", "2006.14480", "2101.06557", "2106.06860", "2106.15004", "2112.01511", "2202.02005", "2203.01387", "2205.15241", "2206.08129", "2207.01206", "2207.14024", "2209.05451", "2210.03094", "2301.04195", "2302.00111", "2302.12766", "2306.02224", "2306.03310", "2307.04577", "2309.02473", "2310.07896", "2310.08864", "2311.02304", "2402.12498", "2403.12945", "2404.06229", "2405.12213", "2406.00960", "2406.09246", "2406.14947", "2406.19464", "2407.01392", "2408.11048", "2409.14411", "2410.11013", "2411.01349", "2411.03817", "2411.04005", "2411.04331", "2411.10291", "2411.12982", "2411.14256", "2411.14770", "2412.04659", "2412.05552", "2412.09858", "2412.11974", "2412.20226", "2501.09747", "2502.05855", "2502.19645", "2503.03734", "2503.14734", "2503.24070", "2507.15062", "2509.17053", "2510.16729", "2512.13100", "2601.07718", "2603.10448", "2605.08713", "2606.20092", "2610.14004"]
related: [learning-from-demonstration, inverse-reinforcement-learning, policy-gradient, actor-critic, dataset-curation, action-tokenization, diffusion-models, continual-learning, robot-simulation, vision-language-action-model, flow-matching, llm-agent, nearest-neighbor-search, self-supervised-learning, masked-image-modeling, transfer-learning, branch-and-bound, graph-neural-network, multimodal-fusion, sim-to-real-transfer, gaussian-splatting, low-rank-adaptation, curriculum-learning, world-model, bev-representation, zero-shot-learning, teleoperation, human-in-the-loop, action-chunking, data-augmentation, vision-language-pretraining, hierarchical-reinforcement-learning, control-barrier-function, model-predictive-control, mixture-of-experts, adaptive-layer-normalization, neural-scaling-laws, knowledge-distillation, topological-map-navigation, domain-randomization, chain-of-thought-prompting, contrastive-learning, offline-reinforcement-learning, object-queries, safety-filter, impedance-control, digital-twin, learned-simulator, motion-prediction]
status: draft
updated_at: 2026-11-01
---

## Overview

Imitation learning fits a policy to examples of what a demonstrator or a
logged system did. The core training target is the recorded action, trajectory
or action token, not a reward. The Orbit paper collects teleoperated
demonstrations and trains behaviour-cloning policies on them
[arxiv:2301.04195 §5.2] [arxiv:2301.04195 tab. 2]. The self-driving dataset
paper trains a planner to reproduce the logged future of the ego vehicle
[arxiv:2006.14480 §4]. The shared idea is a supervised target taken from
recorded actions or trajectories. The shared risk is that a small error moves
the agent into states the data never covered. The driving paper names this as
the accumulation of errors once the vehicle leaves the logged path
[arxiv:2006.14480 §2].

The same supervised idea also appears in three forms the first version of this
page did not cover: a lookup over demonstration frames with no action head
[arxiv:2112.01511 §3.2], a transformer that receives a target return as an input
token [arxiv:2205.15241 §3.1], and an LLM agent that reads a trained imitator's
suggested action from its prompt [arxiv:2306.02224 §2.3]. Imitation also applies outside robotics: a graph
network learns which variable a branch-and-bound solver should split on by
copying an expensive expert rule, strong branching [arxiv:1906.01629 §3.3]
[arxiv:1906.01629 §4.1].

Newer cards add three patterns. Some train diffusion policies on rollouts of a
simulated expert rather than on human recordings. Some learn navigation from
human targets or hand-held cameras. Some use imitation as the supervised
stage of an LLM agent or a language-model-guided driver. Two survey cards
mention imitation without running experiments of their own. One lists it as one
of two end-to-end routes for small research cars [arxiv:2404.06229 §4.4]. The
other groups end-to-end driving stacks under imitation, reinforcement and
language-model approaches, then benchmarks three models on throughput rather
than accuracy [arxiv:2411.10291 §III-B] [arxiv:2411.10291 §V-A].

Several papers on this page add a second signal on top of the recorded data:
RL fine-tuning from a reward [arxiv:2207.01206 §4.3], an adversarial style
reward [arxiv:2601.07718 §3.5], or an inverse-dynamics model that reads actions
from generated frames [arxiv:2302.00111 §3.2]. The reverse also occurs: an RL
learner in a fixed dataset keeps its policy near the logged actions through an
imitation term [arxiv:2106.06860 §5]. A rule-based planner can supply that
recorded action during online RL, too: the parking policy REAP mixes a
behaviour-cloning loss into its SAC actor update [arxiv:2605.08713 §4.2].
Human takeovers are a further source: HACTS records a person's corrections
during a policy rollout and mixes them into the retraining data
[arxiv:2503.24070 §4.2]. HG-DAgger goes further and lets the human decide when
control switches, so the takeover itself is the gate that picks the labelled
states [arxiv:1810.02890 §II]. Two more cards add a model-based expert or a
planner as the teacher: a legged controller cloned by DAgger and then tuned
with PPO [arxiv:2311.02304 §III], and a navigation policy whose goal input can
be masked [arxiv:2310.07896 §IV-A].

Two cards sit at the edges of the page. A survey of offline
RL places imitation among its method classes, as a filter on the logged data
or as a policy conditioned on an outcome [arxiv:2203.01387 tab. 1]. It reports
no experiments of its own. A broader survey of imitation learning sorts the
covariate-shift fixes into methods that need an online expert, methods that
keep the agent near the demonstrated states with a fixed reward, and
constrained methods that avoid unfamiliar states [arxiv:2309.02473 §II]. Its
one table lists 66 works by learning type and online-expert use
[arxiv:2309.02473 tab. I]. Card: [A Survey of Imitation Learning](/papers/2309.02473). A driving transformer trained on a privileged rule
expert adds a rule-based speed controller on top of the learned outputs, and
the authors report that removing it lowers the driving score [arxiv:2207.14024
§3.3] [arxiv:2207.14024 §4.3].

## Variants

### Behaviour cloning on teleoperated demonstrations

Demonstrations are gathered by teleoperation with a keyboard, a gamepad or a
SpaceMouse and stored in the robomimic format [arxiv:2301.04195 §4]
[arxiv:2301.04195 §5.2]. The lift-cube experiment uses 2,000 trajectories for
each of four settings, where the start and goal positions are either fixed or
random [arxiv:2301.04195 §5.2]. Two policies are compared: BC and BC-RNN. The
card does not describe the architecture of BC-RNN [arxiv:2301.04195 tab. 2].
Card: [Orbit](/papers/2301.04195).

Author-reported success over 100 trials per setup [arxiv:2301.04195 tab. 2]:

| Evaluation setup | BC | BC-RNN |
|---|---|---|
| No change | 1.00 | 1.00 |
| Goal changed | 0.89 | 1.00 |
| Initial state changed | 0.47 | 0.88 |
| Goal and initial state changed | 0.43 | 0.87 |

Lifelong manipulation uses the same objective. LIBERO trains each task with
behaviour cloning under a Gaussian-mixture negative log-likelihood, using 50
teleoperated demonstrations per task [arxiv:2306.03310 §4.4]. Its three policy
networks (BC-RNN, BC-T and BC-ViT) all output a Gaussian-mixture action
distribution [arxiv:2306.03310 §4.4]. Card: [LIBERO](/papers/2306.03310).

### Language-conditioned behaviour cloning across many tasks

BC-Z trains one image-to-action policy on 25,877 real-robot demonstrations
covering 100 tasks. Of these, 14,769 were recorded while a learned policy ran
and a person intervened [arxiv:2202.02005 §4]. The task is a sentence turned
into a 512-dim vector by a pretrained sentence encoder, and FiLM layers inject
that vector into a ResNet-18 torso [arxiv:2202.02005 §5.1] [arxiv:2202.02005
§5.3]. Actions are differences from the current state to a pose N steps ahead,
where N is set by an adaptive rule [arxiv:2202.02005 §5.2] [arxiv:2202.02005
app. C]. Author-reported, the language-conditioned policy reaches 32% average
success on held-out tasks, against 4% when the task is given as a human video
[arxiv:2202.02005 tab. 3]. The text says 29 held-out tasks, but the table has
28 rows [arxiv:2202.02005 tab. 2]. The released code covers only the language
route, not the video encoder [Paper vs code of the BC-Z card](/papers/2202.02005).
Card: [BC-Z](/papers/2202.02005).

### Nearest-neighbour lookup on learned features

VINN has no action network. It fine-tunes a ResNet-50 encoder with BYOL on the
demonstration frames, then takes the k closest demonstration frames to a new
image and averages their actions, weighted by a softmin over distance
[arxiv:2112.01511 §3.1] [arxiv:2112.01511 §3.2]. The paper uses k around 10
[arxiv:2112.01511 §4.7]. On a real door-opening task with a Hello Robot Stretch,
the authors report 80% success over 30 rollouts. That compares with 53.3% for
behaviour cloning on frozen representations and 0% for end-to-end behaviour
cloning [arxiv:2112.01511 tab. I]. With the sign and the handle occluded, VINN
keeps 70% against 10% for the frozen-representation baseline. With one bin also
occluded, it keeps 50% against 0% [arxiv:2112.01511 tab. II]. Each query is
compared against every demonstration frame, so the cost grows with the data
[arxiv:2112.01511 §4.8]. Card: [VINN](/papers/2112.01511).

### Behaviour cloning on scripted or simulated experts

VIMA treats a task as a prompt that interleaves text and object images. A
causal decoder predicts discrete pose actions and cross-attends to the prompt.
It is trained by minimising the negative log-likelihood of expert actions over
the steps of each trajectory [arxiv:2210.03094 §4]. The expert trajectories
come from scripted oracles with privileged state [arxiv:2210.03094 §3]. At 200M
controller parameters, the authors report 81.5% on the easiest zero-shot level
against 58.0% for VIMA-Gato, and 48.6% against 16.8% on the hardest level
[arxiv:2210.03094 tab. 10]. With 1% of the imitation data, the level-1 success
falls to 36.3%; with 10% it is 76.3% [arxiv:2210.03094 tab. 11]. The baselines
are the authors' own re-implementations [arxiv:2210.03094 §5.1]. Card:
[VIMA](/papers/2210.03094).

Perceiver-Actor predicts one keyframe action per step over a voxel grid built
from RGB-D input, with a CLIP language goal [arxiv:2209.05451 §3]. Its loss is
four cross-entropy terms, for translation, rotation, gripper state and
collision [arxiv:2209.05451 §3.4]. The paper reports that the single
multi-task agent beats the C2FARM-BC 3D U-Net baseline in 25 of 36
task-and-demo-count cells, with average gains of 1.33× at 10 demonstrations and
2.83× at 100 [arxiv:2209.05451 §4.2] [arxiv:2209.05451 tab. 1]. The keyframe
extraction is essential: the authors report zero success with random or
fixed-interval keyframes [arxiv:2209.05451 §4.2]. Card:
[Perceiver-Actor](/papers/2209.05451).

WebShop uses imitation as the first stage of an agent. A BART model maps the
instruction to a search string, trained on 1,421 instruction–search pairs
[arxiv:2207.01206 §4.2]. A BERT choice model scores the clickable buttons,
trained on 9,558 samples [arxiv:2207.01206 §4.2]. Reported scores on the
500-instruction test split are 59.9 with imitation alone (29.1% success) and
62.4 with imitation followed by policy gradient (28.7% success)
[arxiv:2207.01206 §5.2] [arxiv:2207.01206 tab. 2]. The rule baseline reaches
45.6 and expert humans 82.1 [arxiv:2207.01206 tab. 2]. Reinforcement learning
from a BERT initialisation, without the imitation warm start, performs below
the rule baseline [arxiv:2207.01206 §5.2]. Card: [WebShop](/papers/2207.01206).

### Diffusion-policy imitation with a pretrained visuo-tactile encoder

Touch in the Wild collects demonstrations with a handheld gripper whose soft
fingers carry tactile pads, and a camera films the same moment
[arxiv:2507.15062 §3.1] [arxiv:2507.15062 §3.2]. The encoder is pretrained
first, by reconstructing masked tactile images from the tactile input and the
CLIP embedding of the camera image [arxiv:2507.15062 §4.2] [arxiv:2507.15062
eq. 1]. The policy stage then fine-tunes that encoder inside a diffusion policy
that predicts noise over a 16-step action chunk, sampled with DDIM in 16 steps
[arxiv:2507.15062 §4.3] [arxiv:2507.15062 tab. 5]. On four real-world tasks,
with 20 trials per cell, the test-tube whole-task success is 0.85 with the
pretrained encoder, 0.70 without pretraining and 0.25 for a vision-only policy
[arxiv:2507.15062 tab. 1]. On the pencil whole task the three figures are 0.85,
0.75 and 0.45 [arxiv:2507.15062 tab. 1]. Card:
[Touch in the Wild](/papers/2507.15062).

### Co-training a diffusion policy on a large shared dataset

DROID is a dataset of 76k successful Franka demonstrations, about 350 hours,
across 564 scenes, recorded by many people on one shared hardware stack
[arxiv:2403.12945 §1] [arxiv:2403.12945 §3]. The policy trained on it is a
diffusion policy. Two external views at 128 × 128 pass through ImageNet-
pretrained ResNet-50 encoders, a frozen DistilBERT embeds the instruction, and
a U-Net head denoises 16 action steps, of which 8 run open loop [arxiv:2403.12945
§5.1] [arxiv:2403.12945 app. F.1]. Each training batch is half in-domain
demonstrations and half successful DROID episodes [arxiv:2403.12945 §5.1].
Author-reported, averaged over six lab, office and kitchen tasks, this
co-training adds 22% absolute success in distribution and 17% out of
distribution over the next-best method. Co-training with the Open X-Embodiment
subset is the other comparison [arxiv:2403.12945 §5.2]. Per-task numbers appear
only in a figure. Card: [DROID](/papers/2403.12945).

### Return-conditioned sequence models

Multi-Game Decision Transformer trains one GPT-2-style transformer with
next-token prediction over sequences that hold image patches, a target return,
the action and the reward [arxiv:2205.15241 §3.1] [arxiv:2205.15241 §3.2]. The
training data are offline Atari trajectories from 41 games, with no filtering
between expert and non-expert play [arxiv:2205.15241 §3.3]. At play time the
model samples a high target return before each action [arxiv:2205.15241 §3.4].
Plain behaviour cloning is the same transformer with the return token removed
[arxiv:2205.15241 app. C]. The authors report that the return-conditioned model
beats behaviour cloning on 31 of 41 games [arxiv:2205.15241 §4.7]
[arxiv:2205.15241 fig. 8]. Their headline is an interquartile mean of 126% of
human-normalised score, against 101% for the training data [arxiv:2205.15241 §1].
The 200M-parameter model took 8 days on 64 TPUv4 chips [arxiv:2205.15241
app. B.1]. Card: [Multi-Game Decision Transformers](/papers/2205.15241).

### Imitation term in offline reinforcement learning

TD3+BC keeps a standard deterministic actor-critic algorithm and adds one
imitation term to the actor loss. The actor maximises a Q term scaled by a
weight λ, minus the squared error to the logged action [arxiv:2106.06860 eq. 3].
The weight is set per batch from α = 2.5 and the mean absolute Q value
[arxiv:2106.06860 eq. 5] [arxiv:2106.06860 §5]. The paper states that these
two changes are the only ones made to TD3 [arxiv:2106.06860 §1]. Card:
[A Minimalist Approach to Offline RL](/papers/2106.06860).

On D4RL MuJoCo v0, the authors report a total normalised score of 979.3 over 15
dataset and environment pairs. That compares with 974.6 for Fisher-BRC and
764.3 for CQL [arxiv:2106.06860 tab. 2]. Training takes 39 minutes on one GPU
against 2h 12m for Fisher-BRC and 4h 11m for CQL, using the authors' code
[arxiv:2106.06860 tab. 3]. The gains are not uniform. On HalfCheetah-random,
TD3+BC scores 10.2, below CQL (21.7) and Fisher-BRC (32.2)
[arxiv:2106.06860 tab. 2]. The paper calls the weak random-data results a
concern [arxiv:2106.06860 app. C.1]. The offline RL survey files this method
under policy constraints, and writes the same behaviour-cloning term as its
eq. 20 [arxiv:2203.01387 §IV-A] [arxiv:2203.01387 eq. 20].

### Planner imitation inside an online SAC actor

REAP trains a camera-to-action parking policy with Soft Actor-Critic in a 3D
Gaussian Splatting simulator [arxiv:2605.08713 §4.2] [arxiv:2605.08713 §5.1].
On each training sample, if a Reeds-Shepp planner action exists for the
current state, the actor loss is a mean-squared error to that planner action
with probability p. Otherwise it is the SAC actor loss [arxiv:2605.08713 eq. 4]
[arxiv:2605.08713 eq. 6]. The probability p falls linearly from 1 to 0 over
8,000 epochs [arxiv:2605.08713 §5.3]. The critic sees privileged bird's-eye-view
maps that the deployed actor does not have [arxiv:2605.08713 §4.2]. On standard
slots the authors report 83.5% parking success, against 73.0% for SAC and 37.0%
for the Reeds-Shepp curve planner [arxiv:2605.08713 tab. 1]. The baselines are
adapted to a shared protocol and are not exact reproductions [arxiv:2605.08713
tab. 1]. In the ablation, the row labelled "w/o hybrid RL (stage 2)"
drops to 0.0% success, and removing the perception pretraining gives 50.5%
[arxiv:2605.08713 tab. 2]. The card lists the RL hyperparameters the paper does
not state, including the auxiliary loss weight and the SAC temperature
[arxiv:2605.08713 §4.1] [arxiv:2605.08713 §4.2]. Card:
[REAP](/papers/2605.08713).

### Frozen visual features for small robot imitation heads

Voltron is a representation, not a policy. It is a masked autoencoder over video
frames that either reads the paired caption as a condition or generates it,
chosen per example by a single probability alpha [arxiv:2302.12766 §3] [arxiv:2302.12766 §3.2].
The encoder is then frozen, and a small head is trained on its features for each
robot task [arxiv:2302.12766 §4]. With 25 demonstrations, the language-conditioned
Voltron model reaches 38.2 ± 5.09% success on the imitation task. The variant
without language reaches 33.1 ± 4.79%, and a masked-language baseline reaches
24.0 ± 4.21% [arxiv:2302.12766 tab. 4]. The gap to the no-language variant is
about one standard deviation. Card: [Voltron](/papers/2302.12766).

### Large generalist vision-language-action policies

Octo is an open transformer policy pretrained on about 800k trajectories from
Open X-Embodiment [arxiv:2405.12213 §1] [arxiv:2405.12213 §3.2]. Camera views,
a language or goal-image task and readout slots become tokens, and a small
diffusion head over action chunks decodes the motion [arxiv:2405.12213 §3.1]
[arxiv:2405.12213 §3.3]. A new camera gets a new encoder and a new action space
gets a new head, while the transformer weights are kept [arxiv:2405.12213 §3.1].
Each finetuning setup uses about 100 target trajectories and 50k steps
[arxiv:2405.12213 §3.3]. Author-reported: 72% average success over six
real-robot finetuning setups, against 20% for training from scratch, the best
baseline [arxiv:2405.12213 tab. 1]. Card: [Octo](/papers/2405.12213).

Open X-Embodiment pools 60 existing robot datasets from 34 labs and 22
embodiments into one RLDS format [arxiv:2310.08864 §3.1]. Each dataset's
actions are normalised before they are binned [arxiv:2310.08864 §4.1]. RT-1-X
is trained on the mixture alone, and RT-2-X is co-fine-tuned with web
vision-language data in about a one-to-one split [arxiv:2310.08864 §4.3]. Both
predict action tokens with a categorical cross-entropy loss [arxiv:2310.08864
§4.3]. The mixture weights are not stated [arxiv:2310.08864 §4.3]. The authors
report that RT-1-X beats each dataset's original method on 4 of 5 small-data
domains, but the per-dataset numbers appear only in a bar plot
[arxiv:2310.08864 §5.1] [arxiv:2310.08864 fig. 3]. On the large Bridge
evaluations it does not beat RT-1, while RT-2-X does [arxiv:2310.08864 tab. 1].
RT-2-X raises emergent-skill success from 27.3% to 75.8% against RT-2
[arxiv:2310.08864 tab. 2]. Card: [Open X-Embodiment](/papers/2310.08864).

OpenVLA fine-tunes a 7B vision-language model to predict discretised action
tokens with next-token cross-entropy, computed only on the action tokens
[arxiv:2406.09246 §3.2]. Its pretraining mixture contains about 970k episodes
from the Open X-Embodiment collection [arxiv:2406.09246 §3.3]. On BridgeData V2
it reports 70.6% against 50.6% for RT-2-X [arxiv:2406.09246 app. B.1.3]. On two
Franka tabletop tasks, full fine-tuning reaches 69.7% and LoRA of rank 32
reaches 68.2%, with 97.6M trainable parameters against 7,188.1M
[arxiv:2406.09246 §5.3]. The two results overlap within their standard errors.
The model takes one image and no proprioception, and runs at about 6 Hz on an
RTX 4090 [arxiv:2406.09246 §3.5] [arxiv:2406.09246 §6]. Card:
[OpenVLA](/papers/2406.09246).

FAST changes the target representation rather than the loss. Each action chunk
is passed through a discrete cosine transform, scaled and rounded, then
compressed with byte-pair encoding [arxiv:2501.09747 §5.2]. For a 50 Hz
14-dimensional chunk, the naive per-step binning needs 700 tokens and FAST
needs 53 [arxiv:2501.09747 tab. 1]. The authors report that binning makes no
progress on two real-robot tasks at 20 and 50 Hz [arxiv:2501.09747 §6.2]. The
cost is inference time: about 750 ms per chunk against about 100 ms for
diffusion pi0 on an RTX 4090 [arxiv:2501.09747 §6.5]. Card:
[FAST](/papers/2501.09747).

EventVLA adds a learned memory to a VLA. Fixed visual anchors (the first frame
and recent frames) are combined with a keyframe head that predicts which future
steps will be worth storing [arxiv:2606.20092 §3.1] [arxiv:2606.20092 §3.2]. On
the authors' RoboTwin-MeM benchmark, anchors alone reach 18.0% and anchors with
the keyframe memory reach 75.2% [arxiv:2606.20092 tab. 2]. On RMBench, the
anchors-only variant gives the headline 67.8%, and adding the keyframe memory
gives 66.7% [arxiv:2606.20092 tab. 4]. Inference is 0.42 s per chunk on an A100,
against 1.43 s for MemER [arxiv:2606.20092 tab. 11]. Card:
[EventVLA](/papers/2606.20092).

DiT4DiT starts from a video generator rather than an image-text model. It
pairs a video diffusion transformer with an action diffusion transformer, and
trains both with flow-matching losses [arxiv:2603.10448 §4.3]. The action head
reads hidden features from one video block at a fixed timestep, so control does
not generate frames [arxiv:2603.10448 §4.4] [arxiv:2603.10448 alg. 2]. Trained
from scratch on LIBERO data alone, it reports 98.6% average success, against
97.1% for OpenVLA-OFT and 96.9% for π0.5 [arxiv:2603.10448 tab. 1]. On 24
RoboCasa-GR1 tasks it reports 50.8%, against 41.8% for GR00T-N1.5, which starts
from pretrained weights [arxiv:2603.10448 tab. 2]. It is also the slowest of
the three models in the authors' efficiency table, at 6 Hz against 9 Hz and
13 Hz [arxiv:2603.10448 tab. 3]. Card: [DiT4DiT](/papers/2603.10448).

GR00T N1 is a humanoid VLA. An Eagle-2 backbone reads images and text, and a
diffusion transformer trained with flow matching outputs 16-step action chunks
[arxiv:2503.14734 §2.1]. Sampling uses four Euler steps [arxiv:2503.14734 §2.1].
Training mixes real, simulated, neural-generated and human-video data. Videos
without motor commands get pseudo-actions from latent actions or from an
inverse-dynamics model [arxiv:2503.14734 §2.2] [arxiv:2503.14734 §2.3]. The
author-reported simulation average is 45.0%, against 33.4% for Diffusion Policy
[arxiv:2503.14734 tab. 2]. On real GR-1 tasks with 10% of the post-training
data, the authors report 42.6% against 10.2% for Diffusion Policy
[arxiv:2503.14734 tab. 3]. Card: [GR00T N1](/papers/2503.14734).

DexVLA pairs a Qwen2-VL-2B vision-language model with a roughly 1B-parameter
diffusion-transformer action expert [arxiv:2502.05855 §3.1] [arxiv:2502.05855
app. C.3]. Training has three stages: expert pre-training across embodiments,
alignment to one embodiment, and task post-training [arxiv:2502.05855 §3.2]. The
VLM also writes sub-step text, which the model learns to generate as intermediate
output [arxiv:2502.05855 §3.2]. Author-reported, on their own robots: 0.92 on
shirt folding with no task-specific training [arxiv:2502.05855 §4.1], and 0.4
against 0.2 for pi0 on laundry folding under direct prompting [arxiv:2502.05855
§4.3]. In the expert-size ablation on shirt folding, a 93M UNet scores 0.17, a
410M diffusion expert 0.63 and a 1B expert 0.92 [arxiv:2502.05855 §4.4]. Card:
[DexVLA](/papers/2502.05855).

### Fine-tuning a VLA with parallel decoding and an L1 head

OpenVLA-OFT keeps the OpenVLA-7B pretraining and changes how it is fine-tuned
[arxiv:2502.19645 §3]. Three changes form the recipe: all action slots of a
chunk are decoded in one forward pass with bidirectional attention, a continuous
head replaces the token output layer, and training minimises the mean L1 error
over the chunk [arxiv:2502.19645 §4.2] [arxiv:2502.19645 app. A.1]. The LoRA
rank is 32 [arxiv:2502.19645 tab. 4]. On LIBERO, with a wrist image and
proprioception on the filtered data, the authors report 97.1% average success,
against 76.5% for fine-tuned OpenVLA [arxiv:2502.19645 tab. 1]. Throughput rises
from 4.2 Hz to 108.8 Hz with parallel decoding and chunking, about 26 times
[arxiv:2502.19645 tab. 2]. On a bimanual ALOHA robot at 25 Hz, the FiLM variant
OFT+ scores at least as high as every baseline on all four tasks
[arxiv:2502.19645 tab. 10] [arxiv:2502.19645 tab. 13]. On scooping it scores
100 against 93.33 for pi0, and on the pot task 51.25 against 44.17 for RDT-1B.
Without FiLM, language following falls to chance (33%) on the two
language-dependent tasks [arxiv:2502.19645 §6.3] [arxiv:2502.19645 fig. 5]. Card:
[OpenVLA-OFT](/papers/2502.19645).

### Instruction-aware visual tokens on a frozen CLIP encoder

OTTER keeps CLIP frozen and does not fine-tune it. For each instruction word,
it takes a softmax over the image patches, using the similarity between the
word and each patch, and sums the patches with those weights
[arxiv:2503.03734 §3.1] [arxiv:2503.03734 eq. 4]. The selected tokens are
pooled per camera, and a causal transformer predicts chunks of 10-dimensional
delta end-effector actions [arxiv:2503.03734 §3.2]. The paper states that
only the pooling layers, the proprioception encoder, the transformer and the
action head are trained [arxiv:2503.03734 §3.2]. On 70 unseen real-robot
pick-and-place trials, the authors report 62% for OTTER against 12% for
fine-tuned Octo and 9% for fine-tuned OpenVLA [arxiv:2503.03734 §5.1].
Fine-tuning CLIP itself drops unseen success to 15% [arxiv:2503.03734 §5.3]
[arxiv:2503.03734 app. C]. In simulation, the unseen-task average is 61%
against 48% for fine-tuned OpenVLA [arxiv:2503.03734 §5.2]. Card:
[OTTER](/papers/2503.03734).

### Augmenting logged demonstrations to new robots

OXE-AugE does not collect new demonstrations. It cross-paints existing
Open X-Embodiment frames onto other arms: the source arm is masked, the
background is inpainted, and the trajectory is replayed in simulation for the
target arm [arxiv:2512.13100 §4.1] [arxiv:2512.13100 §4.2]. The pooled set has
16 source datasets and over 4.4 million trajectories, with 8 or 9 target
robots per dataset [arxiv:2512.13100 §6] [arxiv:2512.13100 app. A.2]. The
augmented data is then used to fine-tune OpenVLA-OFT and pi0. The authors
report average gains of 24% and 45% over the no-augmentation baseline on
unseen robot and gripper pairs, in real tests [arxiv:2512.13100 §6.2]. In
simulation, Diffusion Policy trained on the augmented robots is compared on
five Robosuite tasks with 200 demonstrations each [arxiv:2512.13100 §5]. A
diffusion-model version of the same augmentation lowers policy success by 27
to 30% [arxiv:2512.13100 §6.2]. Card: [OXE-AugE](/papers/2512.13100).

### Planning from logged or generated sequences

The driving planner is a ResNet-50 that reads a 224 × 224 bird's-eye-view
raster centred on the ego vehicle and predicts its future (x, y) positions over
five seconds. It uses the same squared-error loss as the forecasting baseline
[arxiv:2006.14480 §4]. The authors follow ChauffeurNet, which augments
imitation training with synthetic perturbations to limit the accumulation of
errors [arxiv:2006.14480 §2] [arxiv:2006.14480 §4]. Evaluation is closed-loop
on 25-second scenes. The network drives the ego vehicle, and the other agents
replay their logs, so they do not react [arxiv:2006.14480 §4]. The paper gives
no numbers for this planning result. It shows a plot and states that
performance rises with the amount of training data [arxiv:2006.14480 §4]
[arxiv:2006.14480 fig. 8]. Card: [One Thousand and One Hours](/papers/2006.14480).

UniPi imitates by generating video. A text-conditioned video diffusion model
produces a plan from the first observed frame and the task text, and an
inverse-dynamics model reads actions from the generated frames
[arxiv:2302.00111 §3.1] [arxiv:2302.00111 §3.2]. The actions are executed
open-loop [arxiv:2302.00111 §3.2]. On seen place instructions in simulation,
the authors report 59.1% against 19.4% for a state-based transformer behaviour
cloner [arxiv:2302.00111 tab. 1]. On the multi-environment suite, the Place
Bowl task gives 51.6% against 9.8% [arxiv:2302.00111 tab. 3]. Pretraining on
internet video raises real-robot success from 72.6% to 77.1%
[arxiv:2302.00111 tab. 4]. Training used 256 TPU-v4 chips [arxiv:2302.00111
app. A.1]. No official code was found (see the card's Paper vs code section). Card:
[UniPi](/papers/2302.00111).

Diffusion Forcing trains a causal sequence model in which every token has its
own noise level, so one model can both predict and plan
[arxiv:2407.01392 §3.1] [arxiv:2407.01392 §3.2]. On the single-task Maze2D
benchmark, the authors report an average reward of 141.7 against 119.5 for
Diffuser* [arxiv:2407.01392 §4.2]. In a robot task that needs memory (swapping
an apple and an orange), the method reaches 80% success, and a diffusion policy
without memory fails [arxiv:2407.01392 §4.4]. The paper warns that the
reward on Maze2D can favour slow plans that stay near the goal
[arxiv:2407.01392 app. D]. Card: [Diffusion Forcing](/papers/2407.01392).

PGP is a motion predictor, not a control policy, but its route sampler is trained
by behaviour cloning. The sampler is a learned discrete policy over a lane graph,
and its loss is the negative log-likelihood of the ground-truth route
[arxiv:2106.15004 §4.2] [arxiv:2106.15004 eq. 3]. The decoder that follows is
trained with a winner-takes-all regression loss over sampled trajectories
[arxiv:2106.15004 §4.3] [arxiv:2106.15004 eq. 5]. On nuScenes, the authors
report MinADE10 of 1.00 and MissRate10,2 of 0.37, against 1.16 and 0.46 for P2T,
the closest prior work [arxiv:2106.15004 tab. 1]. The authors report about 1 minute per
pretraining epoch and 5 minutes per fine-tuning epoch on four V100 GPUs
[arxiv:2106.15004 app. A.7]. No official code is known ([PGP card](/papers/2106.15004)).
Card: [PGP](/papers/2106.15004).

IR-WM trains its planner with an imitation term and a collision term. The
imitation term measures the distance from each planned waypoint to the logged
ego trajectory [arxiv:2510.16729 §3.3] [arxiv:2510.16729 eq. 13]. Its world
model predicts a residual change of the bird's-eye-view state, added to the
previous state, rather than the whole next scene [arxiv:2510.16729 §3.2.2]
[arxiv:2510.16729 eq. 6]. At inference the default planner reads the generated
BEV features, and occupancy is an auxiliary output [arxiv:2510.16729 §3.2.4].
On nuScenes open-loop planning, the authors report an average L2 of 0.53 m
against 0.85 m for Drive-OccWorld, and a collision rate of 0.17% against 0.29%
[arxiv:2510.16729 tab. 4]. SSR reports a lower collision rate, 0.15%
[arxiv:2510.16729 tab. 4]. Card: [IR-WM](/papers/2510.16729).

### Closed-loop unrolled imitation for multi-agent traffic

TrafficSim learns one joint policy for every traffic actor in a scene from
human driving logs. A latent variable shared by the whole scene gives several
plausible joint plans from the same map and history [arxiv:2101.06557 §3.2]
[arxiv:2101.06557 eq. 1]. Training runs the policy in closed loop for 12 s,
with the gradient passing through the sampled actions, so errors that build up
in its own rollout are penalised [arxiv:2101.06557 §3.4]. The imitation term is
a CVAE evidence bound on the first 8 s only, and a collision penalty acts on
prior samples [arxiv:2101.06557 §3.4] [arxiv:2101.06557 app. A]. The weight
shifts from imitation to collision avoidance over time [arxiv:2101.06557 eq. 3].
On ATG4D the scenario collision rate is 0.50%, against 2.90% for the next best
model, ILVM [arxiv:2101.06557 tab. 1]. Traffic-rule violations are 2.77%
against 0.25% for IDM, so the rule metric does not favour TrafficSim
[arxiv:2101.06557 tab. 1]. Training open-loop instead gives a collision rate of
5.92% [arxiv:2101.06557 tab. 2]. The authors report that a planner trained on
simulated traffic reaches a collision rate of 10.73%, against 10.56% for real
data [arxiv:2101.06557 tab. 3]. The printed time-weight formula does not give
the schedule the text describes [arxiv:2101.06557 app. A]. No official code was
found. Card: [TrafficSim](/papers/2101.06557).

### Waypoint and control heads with trajectory-guided attention

TCP is a driving network from one front camera. A shared image encoder feeds
two branches. The trajectory branch predicts waypoints. The control branch
predicts steering, throttle and brake for the current step and for K = 4
future steps, and at each future step it attends over the image grid with
weights from the trajectory branch's hidden state [arxiv:2206.08129 §3.2]
[arxiv:2206.08129 §3.2.2] [arxiv:2206.08129 eq. 2] [arxiv:2206.08129 eq. 3].
Both branches imitate Roach, a reinforcement-learning expert that sees
privileged bird's-eye-view input [arxiv:2206.08129 §3.1]. At test time the two
outputs are blended with weight 0.3 by a turning rule [arxiv:2206.08129 §3.4]
[arxiv:2206.08129 alg. 1]. The authors report a driving score of 69.714 for one
model and 75.137 for the ensemble, in closed-loop CARLA evaluation
[arxiv:2206.08129 tab. 1] [arxiv:2206.08129 §4.1]. The repository's LICENSE is
Apache-2.0, but the paper's supplement says CC BY-NC 4.0, and the code differs
from the paper in several places. Check the card's Paper vs code section before
reusing either. Card: [TCP](/papers/2206.08129).

### Multi-sensor transformer that imitates a privileged expert, with a safety controller

InterFuser sends three RGB views, a focus crop of the front image and a LiDAR
bird's-eye-view histogram into one transformer encoder. Separate decoder
queries then produce waypoints, a 20 × 20 object density map and traffic-rule
outputs [arxiv:2207.14024 §3.1] [arxiv:2207.14024 §3.2]. The training data
come from a rule-based expert with privileged access to the simulator, sampled
at 2 FPS [arxiv:2207.14024 §4.1]. A safety controller reads the density map
and the traffic outputs, sets a desired speed, and stops for red lights and
stop signs [arxiv:2207.14024 §3.3] [arxiv:2207.14024 app. B]. The density map
and traffic outputs are a readable record of what the model perceives, which
the authors present as the interpretable part of the design [arxiv:2207.14024
§1].

The authors report a CARLA leaderboard driving score of 76.18, route
completion of 88.23 and infraction score of 0.84, accessed June 2022
[arxiv:2207.14024 tab. 1]. Removing the safety controller lowers the driving
and infraction scores by 15% and 22%. Concatenating the views instead of
fusing them with the transformer lowers them by 58% and 63% [arxiv:2207.14024
§4.3]. All evaluation is in simulation, and real-road use is left as future
work [arxiv:2207.14024 §5].

The released controller is a fixed rule set, not the linear programme that
appendix B describes, and several loss weights and sizes differ from the
paper's tables. A reimplementation that follows the paper's safety controller
or loss must be written from the paper, not taken from the release. Card:
[InterFuser](/papers/2207.14024).

### Style imitation alongside reinforcement learning

Hiking in the Wild trains a humanoid policy with PPO. Its style term comes from
adversarial motion priors: a discriminator reads short state sequences from
reference motion and the policy is rewarded for looking like it
[arxiv:2601.07718 §3.5]. The style term has weight 0.25 and is trained with a
least-squares loss [arxiv:2601.07718 tab. 5] [arxiv:2601.07718 §3.5]. Walking
and running use separate reference sets, and the running reference is a single
gait cycle of 1.54 s [arxiv:2601.07718 app. A.9]. In simulation, removing the
style term drops success to 0.00% on small-box and ascending-stair terrain,
against 99.09% and 100.0% with it [arxiv:2601.07718 tab. 2]. Card:
[Hiking in the Wild](/papers/2601.07718).

The official code is under CC BY-NC 4.0 ([Hiking card](/papers/2601.07718)),
so do not reuse it commercially. This page uses only the paper's content.

### Recording human corrections during a policy rollout

HACTS is a leader-follower teleoperation arm built from 3D-printed parts and
low-cost servos. Its two arms are synchronised both ways, so the leader follows
the robot while a policy runs, and a person can take over at any point
[arxiv:2503.24070 §3.1] [arxiv:2503.24070 §3.3]. The takeovers are recorded
as correction data and mixed with the original demonstrations for retraining
[arxiv:2503.24070 §4.2]. The authors use the same arm for online RL: a
pretrained actor is fine-tuned with RLPD, with short human corrections
[arxiv:2503.24070 §4.4]. On three real UR5 tasks with ACT, the authors compare HACTS-ACT with a policy
trained on 100 expert demonstrations. In the failure-correction setting, the
UprightMug value is 70% against 40% for that baseline, and OpenBox and
SteamBun favour HACTS-ACT in the same setting [arxiv:2503.24070 tab. 2]. The
two arms of this comparison differ in data: HACTS-ACT gets 50 base
demonstrations plus 50 correction trajectories [arxiv:2503.24070 §4.2].
On CloseBin, RLPD-HACTS goes from 50% to 80% after 45 minutes of online
training, on 10 trials [arxiv:2503.24070 tab. 4]. The paper gives no official
code, so the design is not reproduced here [Paper vs code of the HACTS card](/papers/2503.24070).
Card: [HACTS](/papers/2503.24070).

### Human-gated takeovers with ensemble doubt

HG-DAgger keeps DAgger's loop of training on aggregated data and rolling the
novice out again. The difference is the gate: the human expert decides when to
take control, instead of a coin flip with probability β, and labels are kept
only while the expert is in control [arxiv:1810.02890 §II] [arxiv:1810.02890
eq. 2]. The novice is an ensemble. Its doubt is the l2-norm of the diagonal of
the ensemble's output covariance, and the threshold is the mean of the final
quarter of the doubts logged at takeovers [arxiv:1810.02890 eq. 3] [arxiv:1810.02890
eq. 4]. The threshold is used only to draw a permitted set when the trained
policy is evaluated. Using it as a control gate at test time is left as future
work [arxiv:1810.02890 §IV-B] [arxiv:1810.02890 §V]. In simulation, the authors
report fewer road departures and collisions than DAgger and behaviour cloning
[arxiv:1810.02890 §IV-A]. On a real car with five fixed obstacle
configurations, HG-DAgger has no collisions and no road departures. Behaviour
cloning has one collision and six departures, and DAgger has one of each. The
authors call this test too small for statistical significance
[arxiv:1810.02890 tab. 2] [arxiv:1810.02890 §IV-C]. The method needs a person
who can spot danger and react in real time during training [arxiv:1810.02890
§II]. Card: [HG-DAgger](/papers/1810.02890).

### Demonstrations as teaching

Pedagogical inverse reinforcement learning models a teacher who picks
trajectories that make the true reward easiest for an observer to infer
[arxiv:2610.14004 §2.2] [arxiv:2610.14004 §2.3]. In a goal-grid study with 59
analysed participants, people asked to show the task chose disambiguating final
approaches more often than people asked to do it (1.82 against 1.05)
[arxiv:2610.14004 §3.2]. In a feature-reward study with 58 participants,
showing produced many more revisits per trajectory (1.38 against 0.10)
[arxiv:2610.14004 §4.2.2]. Maximum-likelihood IRL recovers the reward better
from showing trajectories than from doing trajectories, but the paper gives
these distances only in a figure [arxiv:2610.14004 §4.3]. The released code has
no teacher model (see the card's Paper vs code section). Card:
[Showing versus Doing](/papers/2610.14004).

### Expert hints for an LLM agent

An Auto-GPT-style loop lets a chat model choose a command, call a tool, read the
result and choose again. The "additional opinions" method adds a trained
imitation model's proposed action to the prompt. The model still makes its own
choice, and nothing is retrained [arxiv:2306.02224 §2.3] [arxiv:2306.02224 alg. 1].
The WebShop expert is a BART model that writes search queries and a BERT model
that scores clickable options, both trained on demonstrations
[arxiv:2306.02224 §2.1.1]. The ALFWorld expert is a text agent trained with DAgger
[arxiv:2306.02224 §2.1.2]. On 50 WebShop instructions, GPT-4 with five opinions
reaches 0.320 success, against 0.227 for the imitation model alone
[arxiv:2306.02224 tab. 1]. On 134 unseen ALFWorld games, GPT-4 with the opinion
reaches 0.515, against 0.306 for the imitation agent [arxiv:2306.02224 tab. 2].
The effect is not uniform across models. On ALFWorld, the opinion moves GPT-3.5
from 0.075 to 0.030 and Claude from 0.082 to 0.090 [arxiv:2306.02224 tab. 2].
Card: [Auto-GPT for Online Decision Making](/papers/2306.02224).

### Imitating an expert solver's decisions

Learn2branch treats each branching decision as a classification over the
fractional variables. A bipartite graph convolutional network scores every
variable, and a softmax over the fractional candidates is trained with
cross-entropy against the variable that strong branching chose
[arxiv:1906.01629 §4.1] [arxiv:1906.01629 §4.3] [arxiv:1906.01629 eq. 3]. The
expert labels are collected from solver runs on generated instances
[arxiv:1906.01629 suppl. §1]. On set covering, the authors report top-1
imitation accuracy of 65.5% for the graph model, against 57.6% for the best
ranking baseline, SVMrank [arxiv:1906.01629 tab. 1]. Solving time is lower than
SCIP's default reliability pseudocost rule in 11 of the 12 problem and size
settings. That count is made from the time table, not stated by the authors. On
hard set covering the times are 1,489.91 s against 1,677.02 s
[arxiv:1906.01629 tab. 2]. The models are trained on easy instances
only, and the authors say they generalise to larger ones [arxiv:1906.01629
tab. 2] [arxiv:1906.01629 §5.2]. Card:
[Exact Combinatorial Optimization with Graph Convolutional Neural Networks](/papers/1906.01629).

### Cloning a model predictive controller, then RL fine-tuning

The expert is a model predictive controller built on differential dynamic
programming, with friction constraints for the stance legs and Raibert foot
placement for the swing legs. It runs only in simulation, because it cannot run
in real time on the robot [arxiv:2311.02304 §III-A]. A small MLP with layers of
512, 256 and 64 units is cloned from the expert by DAgger on flat terrain. Its
input leaves out body linear velocity, because the authors found that a noisy
state estimator hurt hardware [arxiv:2311.02304 §III-B]. A PPO stage then
fine-tunes the clone on rough, step and cliff terrain, with an asymmetric
actor-critic, a minimal reward and limited exploration [arxiv:2311.02304 §III-C].
The authors report that the fine-tuned policy survives longer than the MPC
expert on rough and slippery ground, and that it tracks commands and keeps gait
symmetry better than vanilla RL [arxiv:2311.02304 §IV-B] [arxiv:2311.02304
§IV-C] [arxiv:2311.02304 §IV-D]. Optimisation time falls from 0.0154 s to 0.0010 s
[arxiv:2311.02304 §IV-B]. On the Mini Cheetah, IFM(CR) crosses a 7.5 cm
step in 9 of 10 trials, and vanilla RL in 8 of 10 [arxiv:2311.02304 tab. 2]. The
hardware trial counts are small, so the authors' differences may not be
meaningful [arxiv:2311.02304 §IV-A]. The card records no official code
[Paper vs code of the card](/papers/2311.02304). Card:
[Imitating and Finetuning Model Predictive Control](/papers/2311.02304).

### Diffusion policies trained on expert rollouts

Perturbation-robust control from a simulated expert. PDP trains one RL expert
per subset of motions, then rolls the experts out with noise on the executed
action while storing the clean expert action as the target [arxiv:2406.00960
§3.2]. A diffusion policy trained on those pairs reaches 100.0% in-distribution
and 96.3% out-of-distribution success on Bump'em perturbation recovery, against
81.0% for an MLP [arxiv:2406.00960 tab. 3]. The paper finds a one-step action
horizon best, and a twelve-step horizon drops to 6.5% in-distribution
[arxiv:2406.00960 tab. 4]. Inference costs K times an MLP's
[arxiv:2406.00960 §6]. Card: [PDP](/papers/2406.00960).

The same idea works for a wheeled robot with a planner as the expert. LiCS
trains a small Transformer on the commands of the FSMT planner and of a
joystick driver, with Gaussian noise on the executed commands and the clean
command as the target [arxiv:2406.14947 §II-A] [arxiv:2406.14947 §IV-A]. On 66
benchmark worlds at 1.4 m/s it reaches 100% success against 99.49% for FSMT
[arxiv:2406.14947 tab. 2]. On 16 harder worlds it reaches 91.67% against 62.50%
[arxiv:2406.14947 tab. 3]. Its hardware run completes two of three cardboard
tracks [arxiv:2406.14947 tab. 5]. Card: [LiCS](/papers/2406.14947).

Legged and humanoid control can take its data from an RL expert too. An
AMP-trained walking policy produces 24 datasets, varying the randomisation and
the size (500K, 2M and 8M transitions) [arxiv:2411.01349 §3] [arxiv:2411.01349
§5]. On flat ground the 500K sets give no stably walking policy, while the 8M
"All" set matches the source policy's success of 1.0 and its tracking of 0.13
against 0.12 [arxiv:2411.01349 §5.1] [arxiv:2411.01349 tab. 3]. The authors
conclude that walking needs much more data than the manipulation tasks where
diffusion policies are usually tested [arxiv:2411.01349 §6]. Card:
[Domain randomization for humanoid diffusion policies](/papers/2411.01349).

Large sets from a specialist's runs. RP1M records about one million
trajectories from one DroQ specialist per piece and trains one policy on subsets
of 12 to 150 songs [arxiv:2408.11048 §4.3] [arxiv:2408.11048 §5]. Diffusion
Policy with a U-Net keeps in-distribution F1 between 0.45 and 0.55 across
subset sizes, while the BC-MLP baseline falls from 0.529 to 0.250
[arxiv:2408.11048 tab. 2]. On 20 unseen songs, the transformer diffusion policy
rises from 0.186 to 0.316 as the subset grows [arxiv:2408.11048 tab. 2]. Card:
[RP1M](/papers/2408.11048).

### Scaling and conditioning transformer diffusion policies

ScaleDP addresses a transformer diffusion policy that stops improving with
depth. DP-T falls from 80.1% to 74.6% MetaWorld success as its layers go from 8
to 14 [arxiv:2409.14411 §1] [arxiv:2409.14411 fig. 1]. ScaleDP replaces the
cross-attention observation path with adaptive layer norm and removes the
causal mask in action self-attention [arxiv:2409.14411 §3.2]. Scaled from 10M to
1B parameters, it reaches 87.50% against 51.25% for DP-T on four Franka tasks,
and 98.33% against 23.33% on three bimanual UR5 tasks, with 20 trials per task
[arxiv:2409.14411 tab. 2] [arxiv:2409.14411 tab. 3]. Removing the causal mask at
the L size lifts the bimanual average from 66.66% to 90.00%
[arxiv:2409.14411 tab. 4]. Card: [ScaleDP](/papers/2409.14411).

### Hierarchical and subgoal-based imitation

HDP splits the policy into two conditional diffusion models. A Guider predicts
the next object contact, and an Actor denoises 16 steps of actions toward it
[arxiv:2411.12982 §3.2] [arxiv:2411.12982 tab. 8]. The Actor loss adds a
Q-value term weighted by η, which the authors set to 0.001 for most tasks
[arxiv:2411.12982 §4.4] [arxiv:2411.12982 tab. 9]. On the real Move-T task,
stage-2 success is 0.35 for Diffusion Policy and 0.65 for HDP with push data,
and 0.45 against 0.75 with grasp data [arxiv:2411.12982 tab. 5]. The
simulation baseline is a smaller Diffusion Policy that the authors re-ran, so
the gains are within the paper only [arxiv:2411.12982 §6.2]. Card:
[HDP](/papers/2411.12982).

FeudalNav uses three levels and no RL. A latent memory map built from
self-supervised features sits above a waypoint network trained on human
point-clicks, and a worker turns waypoints into 15-degree turns and 0.25 m steps
[arxiv:2402.12498 §4.1] [arxiv:2402.12498 §4.2] [arxiv:2402.12498 §3]. The human
dataset has 300 trajectories and about 103,000 frames [arxiv:2402.12498 tab. 1].
On curved paths in Gibson, the stacked variant reaches 40.50% average success
against 23.97% for NRNS without noise [arxiv:2402.12498 tab. 2]. Card:
[FeudalNav](/papers/2402.12498).

TaKSIE generates the next keyframe image from the current view, the instruction
and a GRU embedding of past keyframes [arxiv:2410.11013 §3.4.1]
[arxiv:2410.11013 §3.3.2]. Its subgoal-conditioned policy is a diffusion-based
goal-conditioned behaviour-cloning policy [arxiv:2410.11013 §3.4.2]. Keyframes
come from a progress curve over each demonstration, and regeneration happens at
a cosine similarity of 0.96 or after 20 steps [arxiv:2410.11013 §3.3.1]
[arxiv:2410.11013 §3.3.3]. On CALVIN the average chain length is 3.18 against
2.80 for SuSIE, and on the real robot the success rate is 71.9% against 56.3%
[arxiv:2410.11013 tab. 2] [arxiv:2410.11013 tab. 3]. Card:
[TaKSIE](/papers/2410.11013).

### Navigation learned from demonstrations and human targets

ViDEN collects demonstrations with a hand-held depth camera and no robot. A
four-layer CNN and a four-layer transformer encode the depth image and the
target, and a diffusion model over waypoints conditions on them [arxiv:2412.20226
§II-A] [arxiv:2412.20226 §II-C] [arxiv:2412.20226 §II-D]. On the hardest of
three levels, it reaches 84% success, against 20% for ViNT with a relative
distance goal and 12% for the BC-ConvMLP baseline, over 25 trials per level
[arxiv:2412.20226 tab. I] [arxiv:2412.20226 §III-C]. Card:
[ViDEN](/papers/2412.20226).

AMR is a transformer that predicts base waypoints and camera tilt from frozen
MAE tokens, LiDAR bins and goal tokens. It is trained by behaviour cloning on
500k planned trajectories in simulation, then refined with DAgger
[arxiv:2411.14770 §III-A] [arxiv:2411.14770 §III-B] [arxiv:2411.14770 §IV]. The
authors report a 3 cm and 1 degree median error on unseen objects
[arxiv:2411.14770 §I]. A pose-estimation baseline with a CAD model and a perfect
map completes 89.0% of runs when the object is visible at the start, against
95.9% for AMR [arxiv:2411.14770 fig. 6] [arxiv:2411.14770 §V-A]. Card:
[AMR](/papers/2411.14770).

LiveNet is a neural controller for two robots that imitates a receding-horizon
MPC. It outputs a reference command and learned penalties for a collision
barrier and a liveness barrier, and a QP layer changes only the speed
[arxiv:2412.04659 §4] [arxiv:2412.04659 §5.1]. In simulation it reports no
collisions and no deadlocks in 50 runs per doorway and intersection scenario
[arxiv:2412.04659 tab. 2]. Card: [LiveNet](/papers/2412.04659).

SAME adds a mixture of experts to the visual query of DUET's cross-attention. It
is trained with DAgger on seven VLN datasets [arxiv:2412.05552 §3.2]
[arxiv:2412.05552 §4]. Its R2R validation-unseen success of 76 matches the
authors' multi-task ScaleVLN run [arxiv:2412.05552 tab. 4]. Mixing datasets in
one DAgger batch drops R2R success to 62.36 from 73.69 with sequential batches
[arxiv:2412.05552 tab. 6]. Card: [SAME](/papers/2412.05552).

NoMaD is one Transformer diffusion policy for both exploration and goal
reaching. During training the goal token is masked with probability 0.5. A
masked token is removed from attention, so at test time the same network
explores when the goal is hidden and heads for the goal image when it is shown
[arxiv:2310.07896 §IV-A]. Actions are short 2D sequences sampled from a
conditional diffusion model, which starts from Gaussian noise and runs K
denoising steps [arxiv:2310.07896 §IV-B] [arxiv:2310.07896 §IV-C]. The paper
does not state K. The authors use the policy inside a topological map of
visited places, and the planner details are not in the paper [arxiv:2310.07896
§III]. Author-reported on a real robot: 98% undirected exploration success with
19M parameters, against 77% for the best published baseline with 335M
[arxiv:2310.07896 tab. 1]. In a separate table, a dedicated Diffusion Policy
also reaches 98% undirected, and goal reaching is 92% for both ViNT and NoMaD
[arxiv:2310.07896 tab. 2]. The code is MIT-licensed, but it differs from the
paper in several places, so see the card's Paper vs code section before
reproducing it. Card: [NoMaD](/papers/2310.07896).

### Force-guided imitation with an impedance torque loop

FILIC feeds an estimated six-axis end-effector wrench into an ACT-style
transformer policy. Two RGB streams and the wrench are fused by cross-attention,
and the policy outputs target poses at 25 Hz with action chunking
[arxiv:2509.17053 §4.1]. A joint-space impedance law runs at 2 kHz to turn each
target into torque commands, so the arm yields on contact [arxiv:2509.17053
§4.1] [arxiv:2509.17053 eq. 5]. The wrench comes from joint torques minus the
torque a digital twin predicts for the same state, multiplied by the
pseudo-inverse of the transposed Jacobian, so no wrist force sensor is needed
[arxiv:2509.17053 §4.2] [arxiv:2509.17053 eq. 6]. In simulated peg insertion,
author-reported success is 68.0% with end-effector position only, 80.0% with
joint torque and 90.0% with the estimated force [arxiv:2509.17053 tab. 2]. On
four real tasks, the estimated-force policy reaches 76.6% to 83.3%, against
26.6% to 46.7% for position only [arxiv:2509.17053 tab. 3]. The released code
covers only the wrench estimator [code: README.md#L1-L9] of
[the authors' repository](https://github.com/OpenGHz/mujoco-wrench-estimator),
not the policy or the controller, and the impedance gains and training settings
are not stated [arxiv:2509.17053 §4.1]. Card:
[FILIC](/papers/2509.17053).

### Visuomotor and manipulation policies with extra inputs

RcP replaces the final pooling of a ResNet18 policy. A cross-attention query
built from the joint and gripper state picks the image region that matches the
robot, and a contrastive loss trains it with the L1 action loss
[arxiv:2411.04331 §3.1] [arxiv:2411.04331 §3.3]. In simulated reaching under all
distractors, success is 76.4% against 31.9% for Spatial-Softmax and 8.3% for
R3M [arxiv:2411.04331 tab. 1]. On a real Franka with a self-distractor, it
reaches 92% against 14% [arxiv:2411.04331 §4.3]. Card: [RcP](/papers/2411.04331).

ManiWAV records sound with a contact microphone in a handheld gripper. A
diffusion policy fuses CLIP vision features with audio features, and the
training audio is overlaid with background sounds and robot-motor noise, each
with probability 0.5 [arxiv:2406.19464 §3.1] [arxiv:2406.19464 §3.2]. On
wiping, the authors report 85% success against 40% for vision only, over 20
rollouts [arxiv:2406.19464 §4.2]. Card: [ManiWAV](/papers/2406.19464).

Object-centric dexterous manipulation trains a transformer planner of wrist
poses by behaviour cloning on human motion from ARCTIC, and a PPO finger
controller adds a bounded wrist residual [arxiv:2411.04005 §4.1]
[arxiv:2411.04005 §4.2]. On trained single-object trajectories in simulation,
completion is 83.8% with augmentation and 59.6% without [arxiv:2411.04005 tab. 4].
Real-robot completion ranges from 64.3% to 100% across six objects
[arxiv:2411.04005 tab. 2]. Card:
[Object-centric dexterous manipulation](/papers/2411.04005).

### Generalist policies trained on relabelled or RL-generated rollouts

RLDG trains one RL expert per task with HIL-SERL, rolls out the successful
episodes, and fine-tunes OpenVLA or Octo on them with the action
log-likelihood [arxiv:2412.09858 §3] [arxiv:2412.09858 eq. 2]. The authors
report gains over fine-tuning on the same number of human-demonstration
episodes of 33% for OpenVLA on FMB insertion and 23% on connector insertion,
and 10% and 37% for Octo [arxiv:2412.09858 §4.2] [arxiv:2412.09858 fig. 4]. On
four unseen connectors, OpenVLA succeeds 73 times in 80, against 49 in 80 for
the best single-connector RL policy [arxiv:2412.09858 §4.3]. The RL side takes
one to three hours of real interaction per task and needs a success classifier
[arxiv:2412.09858 app. B]. Card: [RLDG](/papers/2412.09858).

Emma-X fine-tunes OpenVLA on BridgeData V2 demonstrations relabelled by a
vision-language model. Each segment gets a subtask and a justification, and each
frame gets a next-gripper target and a movement plan before its action tokens
[arxiv:2412.11974 §3.2] [arxiv:2412.11974 §3.3]. Over 12 real WidowX tasks with
10 trials each, the authors report 57.5% success against 33.3% for OpenVLA
[arxiv:2412.11974 tab. 1] [arxiv:2412.11974 §4.2]. Card:
[Emma-X](/papers/2412.11974).

### Imitation for LLM agents and LLM-guided control

StepAgent pairs each expert step with the agent's own action from the same
prefix. The pair becomes an implicit preference in a DPO-style loss, or it
trains a discriminator whose output is a step-wise reward for PPO
[arxiv:2411.03817 §4.1] [arxiv:2411.03817 §4.2]. The main results use only this
step-wise reward, with no final environment reward and no human step labels
[arxiv:2411.03817 §7.2]. On WebShop with Llama-3-8B, the inverse-RL variant
scores 67.6 against 65.7 for the best baseline, ETO [arxiv:2411.03817 tab. 2].
Card: [StepAgent](/papers/2411.03817).

An imitation policy with one head per route class (LEFT, MIDDLE, RIGHT) drives
an RC car, and a frozen multimodal LLM answers with chain-of-thought at its own
pace [arxiv:2411.14256 §2.1] [arxiv:2411.14256 §2.2]. The car keeps the last
answer until a new one arrives. Trained on one cone, the policy alone succeeds
40% of the time on a harder course, against 63% to 83% with an LLM hint,
depending on the model [arxiv:2411.14256 tab. 3]. Card:
[LLM-guided driving](/papers/2411.14256).

## When to use which

| Situation | Use | Why |
|---|---|---|
| One task, a few hundred teleoperated demos, changing starts or goals | Behaviour cloning with a recurrent head (BC-RNN) | BC-RNN reaches 0.88 with a changed initial state against 0.47 for BC [arxiv:2301.04195 tab. 2] |
| Multi-task manipulation with RGB-D cameras and few demos | Voxel keyframe behaviour cloning (Perceiver-Actor) | Average gain over C2FARM-BC of 1.33× at 10 demos [arxiv:2209.05451 §4.2]. Training took 16 days on 8 V100 GPUs [arxiv:2209.05451 §3.4] |
| Tasks given as text plus object images, held-out tests | Prompt-conditioned behaviour cloning (VIMA) | 81.5% against 58.0% at the easiest zero-shot level [arxiv:2210.03094 tab. 10]. Depends on an object detector [arxiv:2210.03094 §5.3] |
| New robot, one GPU, single camera, no proprioception needed | Fine-tune a generalist VLA with LoRA (OpenVLA) | LoRA rank 32 gives 68.2% with 97.6M trainable parameters [arxiv:2406.09246 §5.3] |
| Fast, smooth, dexterous actions where per-step tokens stall | Compressed action tokens (FAST) | 700 naive tokens against 53 for one 50 Hz chunk [arxiv:2501.09747 tab. 1]. Slower inference [arxiv:2501.09747 §6.5] |
| Task hinges on a cue seen briefly mid-episode | Keyframe memory on top of visual anchors (EventVLA) | Anchors alone reach 18.0% on RoboTwin-MeM against 75.2% with the memory [arxiv:2606.20092 tab. 2] |
| Goal changes often, plans should be inspectable before execution | Video planning with an inverse-dynamics model (UniPi) | 59.1% against 19.4% on seen place instructions [arxiv:2302.00111 tab. 1]. Needs video-scale training [arxiv:2302.00111 app. A.1] |
| Legged or humanoid locomotion where the gait must look natural | RL with an adversarial style reward (AMP) | Without the style term, success on ascending stairs drops to 0.00% [arxiv:2601.07718 tab. 2] |
| Lifelong learning across a task sequence | Replay or PackNet on top of behaviour cloning | PackNet has the best AUC on LIBERO-Object and LIBERO-Goal [arxiv:2306.03310 tab. 2] |
| Cheap demonstrations across many robot setups | Vision-based teleoperation with retargeting (AnyTeleop) | Higher imitation success than the baseline on 5 of 6 simulated rows [arxiv:2307.04577 tab. 3] |
| Recovering a reward from deliberate demonstrations | Pedagogical IRL | Better reward recovery from showing trajectories, shown in a figure only [arxiv:2610.14004 §4.3] |
| Driving from logs with perturbations | Imitation of the logged ego future with synthetic perturbations | Perturbations target error accumulation [arxiv:2006.14480 §2]. The card has no closed-loop comparison with and without them |
| Few demonstrations, visual changes at test time, no action network to train | Nearest-neighbour lookup on a BYOL-tuned encoder (VINN) | 70% against 10% under sign and handle occlusion [arxiv:2112.01511 tab. II]. Each query scans all demonstration frames [arxiv:2112.01511 §4.8] |
| Discrete actions, mixed-quality logs, a large compute budget | Return-conditioned transformer (Multi-Game DT) | Beats behaviour cloning on 31 of 41 games [arxiv:2205.15241 §4.7]. 200M model took 8 days on 64 TPUv4 [arxiv:2205.15241 app. B.1] |
| Little robot data, plenty of uncaptioned or captioned video | Pretrain a visual encoder, then a small imitation head (Voltron) | 38.2% against 33.1% without language at 25 demos, a gap of about one standard deviation [arxiv:2302.12766 tab. 4] |
| Strong VLA baseline on a benchmark with its own training set | Video-backed action DiT trained from scratch (DiT4DiT) | 98.6% on LIBERO against 97.1% for OpenVLA-OFT [arxiv:2603.10448 tab. 1]. Control at 6 Hz [arxiv:2603.10448 tab. 3] |
| Driving with a lane map, several plausible futures to score | Route policy trained by behaviour cloning, then decoded trajectories (PGP) | MinADE10 1.00 against 1.16 for P2T on nuScenes [arxiv:2106.15004 tab. 1] |
| An LLM agent already runs, and a trained imitator exists for the same environment | Expert suggestion in the prompt (additional opinions) | 0.320 against 0.227 on WebShop with GPT-4 [arxiv:2306.02224 tab. 1]. Lowers GPT-3.5 on ALFWorld [arxiv:2306.02224 tab. 2] |
| A simulator for RL, and a rule planner that can produce feasible actions for the states visited | SAC with planner behaviour cloning mixed into the actor loss (REAP) | 83.5% against 73.0% for SAC on standard slots, author-reported with adapted baselines [arxiv:2605.08713 tab. 1]. Needs planner queries during training [arxiv:2605.08713 §6] |
| A fixed logged dataset with rewards, and no environment access during training | Actor-critic with a behaviour-cloning term (TD3+BC) | 979.3 total against 764.3 for CQL [arxiv:2106.06860 tab. 2]. Weak on random-quality data [arxiv:2106.06860 tab. 2] |
| Many robot datasets with different embodiments, and small data for the target robot | Co-training on pooled RLDS data (RT-1-X) | RT-1-X beats the original method on 4 of 5 small-data domains [arxiv:2310.08864 §5.1]. It does not beat RT-1 on the large Bridge tasks [arxiv:2310.08864 tab. 1] |
| Contact-rich manipulation with a handheld gripper, data collected in the wild, touch available | Diffusion policy on a tactile-pretrained encoder (Touch in the Wild) | Test-tube whole task 0.85 with pretraining, 0.70 without, 0.25 vision-only [arxiv:2507.15062 tab. 1]. Only 20 trials per cell [arxiv:2507.15062 tab. 1] |
| Branching or other solver decisions on one problem family, where the expert rule is too slow to run at every node | Graph network that imitates strong branching (learn2branch) | Hard set covering 1,489.91 s against 1,677.02 s for SCIP's default [arxiv:1906.01629 tab. 2]. Trained on easy instances only [arxiv:1906.01629 tab. 2] |
| New robot with about 100 target demonstrations, a new sensor or action space, and open weights wanted | Fine-tune a transformer generalist with a diffusion action head (Octo) | 72% against 20% for training from scratch, over six setups with 20 trials each [arxiv:2405.12213 tab. 1]. Each setup uses about 100 target trajectories [arxiv:2405.12213 §3.3] |
| Fine-tuning a 7B VLA where one token per step is too slow for the control rate | Parallel decoding with an L1 head on action chunks (OFT) | 108.8 Hz against 4.2 Hz for OpenVLA's decoding [arxiv:2502.19645 tab. 2]. Chunks run open-loop, 8 steps on LIBERO [arxiv:2502.19645 §5.1] |
| Long, multi-stage tasks where the model should narrate sub-steps, and budget exists to annotate them | Staged training with sub-step text and a diffusion expert (DexVLA) | Laundry folding 0.4 against 0.2 for pi0 [arxiv:2502.05855 §4.3]. Without stage 3 the laundry score falls to 0 [arxiv:2502.05855 §4.4]. Sub-step labels are not released |
| Humanoid with little teleoperated data, and human or generated video available | Co-train on real, simulated and video-derived data, then fine-tune a flow-matching action head (GR00T N1) | 42.6% against 10.2% for Diffusion Policy with 10% of post-training data [arxiv:2503.14734 tab. 3]. Neural trajectories add 5.8 points over 8 GR-1 tasks [arxiv:2503.14734 §4.4] |
| Driving from camera images, where the model must also forecast the scene and the planner learns from logged trajectories | Residual BEV world model with a trajectory imitation loss (IR-WM) | L2 0.53 m against 0.85 m for Drive-OccWorld on nuScenes open-loop planning [arxiv:2510.16729 tab. 4]. The fully decoupled planner is faster, 558 ms against 706 ms on a device not stated, but reaches L2 0.87 against 0.53 [arxiv:2510.16729 tab. 5] |
| Many tasks on one robot, a language command at test time, and tasks unseen in training | Language-conditioned behaviour cloning with a FiLM-conditioned policy (BC-Z) | 32% held-out success with language against 4% with video [arxiv:2202.02005 tab. 3]. The video route is not in the release [BC-Z card, Paper vs code](/papers/2202.02005) |
| Few in-domain demonstrations per task, and a large shared dataset from the same hardware | Diffusion policy co-trained 50/50 with the shared data (DROID) | +22% in distribution and +17% out of distribution over the next-best method, author-reported and averaged over six tasks [arxiv:2403.12945 §5.2]. Ten rollouts per task and method [arxiv:2403.12945 §5.2] |
| Keep a large VLM frozen and generalise to new objects with a small robot dataset | Frozen CLIP with instruction-aware visual tokens (OTTER) | 62% against 12% for fine-tuned Octo on 70 unseen pick-and-place trials [arxiv:2503.03734 §5.1]. Fine-tuning CLIP gives 15% [arxiv:2503.03734 §5.3] |
| Robot data concentrated on a few arms, and a new arm or gripper needed without new demonstrations | Cross-painted augmentation of logged data, then fine-tuning (OXE-AugE) | +24% (OpenVLA-OFT) and +45% (pi0) on average over no augmentation, in partial scores on unseen pairs [arxiv:2512.13100 §6.2]. Needs a simulator and a per-robot model |
| A person must take over a running policy, the robot cannot sense force, and corrections should feed training | Teleoperation with bilateral sync and recorded corrections (HACTS) | UprightMug 70% against 40% for full-ACT on the failure-correction setting [arxiv:2503.24070 tab. 2]. Only 10 rollouts per cell [arxiv:2503.24070 §4.2] |
| Simulated expert with state access, contact-rich control, no human demonstrations | Diffusion policy on noise-augmented expert rollouts (PDP) | 96.3% against 81.0% for an MLP out of distribution on Bump'em [arxiv:2406.00960 tab. 3]. A one-step horizon was best [arxiv:2406.00960 tab. 4] |
| Planner available as an expert, wheeled robot with a 2D LiDAR, real-time output needed | Small Transformer cloned from the planner with noisy executed commands (LiCS) | 91.67% against 62.50% for FSMT on 16 hard worlds [arxiv:2406.14947 tab. 3]. Two of three hardware tracks [arxiv:2406.14947 tab. 5] |
| Legged or humanoid control where RL data is cheap to generate | Diffusion policy on expert rollouts, sized by data volume (humanoid study) | 500K transitions gave no stable walk on flat ground; 8M matched the source policy's success [arxiv:2411.01349 §5.1] [arxiv:2411.01349 tab. 3] |
| Transformer diffusion policy that stops improving with size | Adaptive layer norm and non-causal action attention (ScaleDP) | 87.50% against 51.25% for DP-T on four Franka tasks, author-reported [arxiv:2409.14411 tab. 2] |
| Multi-contact task where one diffusion model fails to order the contacts | Contact-guided two-level diffusion with a Q term (HDP) | Move-T stage-2 success 0.65 against 0.35 for Diffusion Policy on push data [arxiv:2411.12982 tab. 5] |
| Visual demonstrations, and the target or subgoal changes during the task | Progress-selected keyframe subgoals with a subgoal policy (TaKSIE) | CALVIN average chain length 3.18 against 2.80 for SuSIE [arxiv:2410.11013 tab. 2] |
| Precise insertion where RL reaches high success in the real world, and a generalist should keep its breadth | Distil successful RL rollouts into a VLA (RLDG) | 73 of 80 on unseen connectors, against 49 of 80 for the best single-connector RL policy [arxiv:2412.09858 §4.3] |
| Long-horizon VLA task where a written plan before each action is affordable | Segment-labelled subtask, reasoning and movement plan before actions (Emma-X) | 57.5% against 33.3% for OpenVLA over 12 real tasks, author-reported [arxiv:2412.11974 tab. 1] |
| Navigation to a target with a hand-held camera and no robot data | Diffusion policy on hand-held depth-camera demonstrations (ViDEN) | 84% against 20% for ViNT with a relative-distance goal on the hardest level [arxiv:2412.20226 tab. I] |
| Goal given as an image, no odometry or metric map, and no RL | Three-level manager and worker trained on human clicks (FeudalNav) | Curved-path success 40.50% against 23.97% for NRNS without noise [arxiv:2402.12498 tab. 2] |
| Precise pose relative to an object, with no CAD model or map | Transformer trained by behaviour cloning, refined with DAgger (AMR) | 3 cm and 1 degree median on unseen objects, author-reported [arxiv:2411.14770 §I] |
| Two robots must pass a doorway or crossing without deadlock and change the path as little as possible | Imitate a receding-horizon MPC, add a QP safety layer (LiveNet) | No collisions and no deadlocks in 50 runs per scenario, in simulation [arxiv:2412.04659 tab. 2] |
| Visuomotor reaching under distractors, including a second robot of the same type | Robot-centric pooling trained with a contrastive loss (RcP) | 76.4% against 31.9% for Spatial-Softmax under distractors [arxiv:2411.04331 tab. 1] |
| Contact sound is informative and the camera misses the contact | Diffusion policy with contact-audio features and noise augmentation (ManiWAV) | Wiping 85% against 40% for vision only [arxiv:2406.19464 §4.2] |
| Dexterous hands follow an object path and human motion data exists | Behaviour-cloned wrist planner with a PPO finger controller (object-centric dexterous) | 83.8% against 59.6% without augmentation on trained trajectories [arxiv:2411.04005 tab. 4] |
| Many VLN tasks with different instruction styles, one model wanted | Mixture of experts on the visual query, trained by DAgger (SAME) | R2R validation-unseen success 76, matching the authors' multi-task ScaleVLN run [arxiv:2412.05552 tab. 4] |
| LLM agent with an SFT base and step-level expert data, no final-reward rollouts | Step-wise reward from expert and agent actions at the same prefix (StepAgent) | WebShop 67.6 against 65.7 for ETO on Llama-3-8B [arxiv:2411.03817 tab. 2] |
| Driving from one simple training layout, real car, LLM available for hints | Imitation policy with a frozen multimodal LLM hint (LLM-guided driving) | 40% alone against 63% to 83% with an LLM on the harder course [arxiv:2411.14256 tab. 3] |
| A human can see danger and take over at any time during training, and an ensemble is affordable for the novice | Human-gated takeovers, labels only during expert control (HG-DAgger) | 0 collisions and 0 road departures on five fixed real-car scenarios, against 1 and 6 for behaviour cloning [arxiv:1810.02890 tab. 2]. Small test with no variance reported [arxiv:1810.02890 §IV-C] |
| Legged robot where a model-based controller works in simulation but cannot run on the robot, and a light policy is wanted | Clone the MPC with DAgger, then fine-tune with PPO (IFM) | Optimisation time 0.0154 s against 0.0010 s [arxiv:2311.02304 §IV-B]. 9 of 10 against 8 of 10 for vanilla RL on a 7.5 cm step [arxiv:2311.02304 tab. 2] |
| Driving from one front camera, where waypoints and direct controls are both wanted | Shared encoder with waypoint and multi-step control branches (TCP) | Driving score 69.714 for one model, 75.137 for the ensemble, in CARLA [arxiv:2206.08129 tab. 1]. The code licence and the supplement's licence disagree, so check both before commercial reuse |
| Driving from several cameras and LiDAR, where a readable object map and a speed limit from that map are wanted, and a privileged rule expert can label the logs | One transformer over all sensor tokens, imitating the expert, with a rule-based speed controller (InterFuser) | Driving score 76.18 on the CARLA leaderboard, author-reported [arxiv:2207.14024 tab. 1]. Removing the safety controller lowers the driving score by 15% [arxiv:2207.14024 §4.3]. The released controller is not the paper's, so write the paper's version yourself [arxiv:2207.14024 app. B] |
| Real robot that must explore unseen places and also reach a goal image, with a small model | One goal-masked diffusion policy (NoMaD) | 98% exploration success with 19M parameters, against 77% for a 335M baseline [arxiv:2310.07896 tab. 1]. Goal reaching 92%, the same as ViNT [arxiv:2310.07896 tab. 2] |
| Simulated road users must react to each other over a long rollout, and logs are the only source of behaviour | Joint latent-plan policy unrolled in closed loop with a fading imitation term and a collision penalty (TrafficSim) | Scenario collision rate 0.50% against 2.90% for ILVM on ATG4D [arxiv:2101.06557 tab. 1]. Open-loop training gives 5.92% [arxiv:2101.06557 tab. 2]. Traffic-rule violations are higher than IDM's [arxiv:2101.06557 tab. 1] |
| Contact-rich insertion or pressing on a low-cost arm with no force sensor, and torque-level control is available | Policy that outputs targets, plus a joint impedance loop fed by a digital-twin wrench estimate (FILIC) | Simulated peg insertion 90.0% against 68.0% with position only [arxiv:2509.17053 tab. 2]. Real tasks 76.6% to 83.3% against 26.6% to 46.7% [arxiv:2509.17053 tab. 3]. Only the estimator is released |
| Need an overview of imitation families before choosing a method | Read a survey, not a method (Zare et al.) | Covers BC, IRL, adversarial IL and imitation from observation [arxiv:2309.02473 §I]. Runs no experiments, and its text and table disagree on some entries [arxiv:2309.02473 tab. I] |
| Demonstration budget is small | Not stated | The cards test 2,000 trajectories per setting (Orbit) and 50 per task (LIBERO); no smaller budget is reported for the same tasks |

The BC results come from one task in one simulator and one robot, so they do
not show that BC-RNN wins on other tasks [arxiv:2301.04195 tab. 2]. The
planner results come from one dataset collected on one route, and the
closed-loop result is qualitative [arxiv:2006.14480 §1] [arxiv:2006.14480 §4].

## Implementation sketch

Common core: a supervised loss on recorded pairs, with the variants changing
the data, the input and the target. Pseudocode lines are our own.

```pseudocode
# Behaviour cloning on recorded (observation, action) pairs

def train_bc(demos, policy, optimiser, steps):
    # Demonstrations come from teleoperation
    # [arxiv:2301.04195 §5.2]
    for step in range(steps):
        obs, action = sample_pair(demos)
        # Loss and step count are not stated on the cards
        loss = action_loss(policy(obs), action)
        optimiser.step(loss)
    return policy


# Planning from logs: perturb the ego state, keep the logged target
# [arxiv:2006.14480 §4]

def perturbed_example(history, future, prob):
    # With probability prob, shift the ego state by random offsets
    if random_uniform() >= prob:
        return history, future
    offsets = gaussian_offsets()
    # The paper gives no offset magnitudes or refit method [§4]
    return history, refit_future(history, future, offsets)
```

Where variants differ: BC maps one observation to one action, while the
planner maps a raster to a sequence of positions. VLA variants (OpenVLA, FAST)
replace the action regression with a cross-entropy over action tokens
[arxiv:2406.09246 §3.2] [arxiv:2501.09747 §5.2]. UniPi replaces the policy
with a video model and reads actions from frames [arxiv:2302.00111 §3.2]. BC-RNN
is the variant with the history dependence shown in the table above. VINN has
no action head: it replaces the regression with a softmin-weighted average over
the k nearest demonstration actions [arxiv:2112.01511 §3.2]. Multi-Game DT keeps
the regression but adds a return token as an input [arxiv:2205.15241 §3.1].
TD3+BC keeps an RL actor objective and adds the squared action error to it
[arxiv:2106.06860 eq. 3]. Learn2branch replaces the regression with a softmax
cross-entropy over candidate variables [arxiv:1906.01629 §4.3]. Touch in the
Wild replaces it with noise prediction in a diffusion policy
[arxiv:2507.15062 §4.3]. Octo uses a diffusion head over the chunk, with 20
denoising steps [arxiv:2405.12213 §3.3] [arxiv:2405.12213 app. D]. OFT replaces
the regression with a single L1 loss over the whole chunk [arxiv:2502.19645
§4.2]. GR00T N1 regresses a flow-matching velocity and integrates it in four
Euler steps [arxiv:2503.14734 §2.1]. DexVLA's expert is a diffusion transformer
whose noise schedule and loss form the paper does not state
[arxiv:2502.05855 §3.1]. The IR-WM planner keeps a trajectory regression and
adds a collision term [arxiv:2510.16729 eq. 13].
Voltron changes only the encoder; the task head is a small network trained on
frozen features [arxiv:2302.12766 §4]. The additional-opinions agent keeps the LLM and
adds the imitator's output as text [arxiv:2306.02224 §2.3]. REAP keeps a SAC
actor and, for each sample, swaps in an MSE loss to the planner action with
probability p [arxiv:2605.08713 eq. 4].

Two design choices recur in the new cards. PDP and LiCS keep the clean expert
action as the target and add noise only to the executed action, which widens
the states the policy visits [arxiv:2406.00960 §3.2] [arxiv:2406.14947 §II-A].
HDP's Actor adds a weighted Q-value term to the denoising loss, with the Q-value
taken on the denoised action [arxiv:2411.12982 eq. 9].

## Contradictions

- **Pretraining.** LIBERO reports that basic supervised pretraining can hurt
  downstream lifelong learning, though the result is shown only in a figure
  [arxiv:2306.03310 §5.2] [arxiv:2306.03310 fig. 5]. OpenVLA reports that its
  pretraining on Open X-Embodiment matters more than its fused encoder: 45.6%
  when trained only on BridgeData V2 against 76.3% with the full mixture
  [arxiv:2406.09246 app. D.1]. OFT's ablation points the same way for robot
  pretraining: 97.1 against 91.9 on LIBERO for a non-robot base
  [arxiv:2502.19645 tab. 15]. The two studies differ in model, data and metric,
  so their results should not be merged.
- **Continuous action head.** Octo reports that its diffusion head beats a
  continuous MSE head (83% against 35% aggregate) and a discretized head (18%)
  in its WidowX ablation [arxiv:2405.12213 tab. 2]. OFT chooses L1 regression.
  Its diffusion variant, run with 50 DDIM steps, appears only as an ablation,
  and its efficiency table reports lower success at fewer steps
  [arxiv:2502.19645 tab. 1] [arxiv:2502.19645 tab. 2]. The setups and the
  alternatives differ, so the two papers do not settle which head to use.
- **pi0 on LIBERO.** DexVLA's table gives pi0 an average of 97.1
  [arxiv:2502.05855 app. A.2]. OFT's table gives fine-tuned pi0 on filtered data
  94.2 [arxiv:2502.19645 tab. 1]. The DexVLA card notes that the paper does not
  say whether its baselines were re-run or copied, so the two figures may come
  from different setups.
- **Per-step binning at low control rates.** OpenVLA uses 256-bin per-dimension
  tokens at 5 Hz and 15 Hz and reports good success on its tasks
  [arxiv:2406.09246 §5.2]. FAST reports that the same binning makes no progress
  on tasks at 20 Hz and 50 Hz [arxiv:2501.09747 §6.2]. Neither paper tests the
  other's control rate or robot, so the findings are not directly in conflict.
- **Architecture ranking.** LIBERO finds that BC-T and BC-ViT beat BC-RNN on
  average, and that the best architecture depends on the algorithm and the
  suite [arxiv:2306.03310 §5.2]. Orbit compares only BC and BC-RNN, on a
  different benchmark [arxiv:2301.04195 tab. 2].

- **Horizon length.** PDP finds a one-step horizon best for perturbation
  recovery and reports near-zero success at twelve steps [arxiv:2406.00960 tab.
  4]. ViDEN executes 16 of 32 predicted steps and reports 84% on its hardest
  navigation level [arxiv:2412.20226 §III-A] [arxiv:2412.20226 tab. I]. The
  tasks, action spaces and controllers differ, so the two results do not settle
  a horizon.
- **Expert versus human data.** RLDG reports that successful RL-expert rollouts
  beat the same number of human demonstrations for OpenVLA and Octo on precise
  insertion [arxiv:2412.09858 §4.2]. LiCS's joystick-driven variant reaches
  98.02% against 100% for its FSMT-expert variant in simulation [arxiv:2406.14947
  tab. 4]. Both favour the non-human source on their own tasks, but the settings
  and margins differ, and PDP and the humanoid study report no human-data
  comparison [arxiv:2406.00960 §5] [arxiv:2411.01349 §5].
- **Baseline from a different codebase.** HDP says the published Diffusion
  Policy numbers could not be matched in its setup, and it re-runs a smaller
  baseline [arxiv:2411.12982 §6.2]. Its within-paper gains do not compare with
  published numbers, which are not on this page.
- **Diffusion-policy noise within one paper.** PDP's text gives 66.9% for noisy
  states with noisy actions on perturbation recovery, while its table gives
  59.5% [arxiv:2406.00960 §5.1] [arxiv:2406.00960 tab. 1]. Quote the table value
  and say so.

## Pitfalls

- **Compounding error.** Open-loop execution is used for the video planner in
  all experiments [arxiv:2302.00111 §3.2], and the single-image VLA takes no
  history [arxiv:2406.09246 §6]. Plan for drift when reproducing either. OFT
  also runs each chunk open-loop, 8 steps on LIBERO and 25 on ALOHA
  [arxiv:2502.19645 §5.1] [arxiv:2502.19645 tab. 5].
- **Benchmarks are not comparable.** WebShop scores partial credit against a
  text-match reward [arxiv:2207.01206 §3.1]. The Maze2D reward can favour slow
  agents [arxiv:2407.01392 app. D]. LIBERO's released code picks each task's
  checkpoint on the same episodes it reports, so its success is optimistic
  ([LIBERO card, Pitfalls](/papers/2306.03310)). LIBERO averages also differ
  by protocol: DexVLA's 97.3 uses stage-1 weights alone [arxiv:2502.05855 app.
  A.2], while OFT's 97.1 fine-tunes each suite and picks the best checkpoint
  [arxiv:2502.19645 tab. 1]. Compare numbers only under the same protocol.
- **Headline numbers differ within one paper.** WebShop's abstract gives a
  29% success rate, while its tables give 29.1% for imitation alone and 28.7%
  for imitation plus RL [arxiv:2207.01206 §5.2] [arxiv:2207.01206 tab. 2]. Name
  the model you quote.
- **Small trial counts.** Real-robot results use 5 to 10 episodes per task in
  Perceiver-Actor [arxiv:2209.05451 tab. 2], 5 to 12 in OpenVLA
  [arxiv:2406.09246 app. B.3.1], and 10 trials per terrain-gait pair in Hiking
  [arxiv:2601.07718 §4.1]. Touch in the Wild uses 20 trials per cell, so one
  trial moves a rate by 0.05 [arxiv:2507.15062 tab. 1]. DexVLA uses 10 trials
  per task and 3 per visual-generalisation setting [arxiv:2502.05855 §4.1]
  [arxiv:2502.05855 app. A.1]. GR00T N1 uses 10 per real task and 5 for one
  task [arxiv:2503.14734 §4.2]. Differences of one or two trials are noise. DROID compares 10 A/B rollouts
per task and method [arxiv:2403.12945 §5.2].
- **Best-checkpoint reporting.** OFT evaluates a checkpoint every 50K steps and
  reports the best, on the same runs it reports [arxiv:2502.19645 §5.1]
  [arxiv:2502.19645 app. A.4]. GR00T N1 takes the maximum over its last five
  simulation checkpoints [arxiv:2503.14734 §4.3]. Both figures are optimistic
  against a fixed-step protocol.
- **Averages do not follow from the columns.** GR00T N1's 45.0% simulation
  average is not the mean of its three suites (32.1, 66.5 and 50.0 give 49.5),
  and its DexMG value changes between tables, 66.5% in tab. 2 and 58.5% in
  tab. 4 [arxiv:2503.14734 tab. 2] [arxiv:2503.14734 tab. 4]. Name the table
  and the suite when you quote a figure.
- **Flow-matching sign convention.** GR00T N1's text defines the interpolation
  as tau times the action plus (1 − tau) times noise, with the target noise
  minus the action and an Euler update with a plus sign [arxiv:2503.14734 §2.1]
  [arxiv:2503.14734 eq. 1]. Taken literally, these do not integrate from noise
  to data consistently. A later release of the authors' code uses the opposite
  sign, which is self-consistent. Check the convention with a toy example before
  training ([GR00T N1 card, Pitfalls](/papers/2503.14734)).
- **Wrong release in the linked repository.** The repository linked from GR00T
  N1 holds a later release with a different backbone and an action horizon of
  40, not 16 ([GR00T N1 card, Paper vs code](/papers/2503.14734)). Pin the N1
  release or implement from the paper.
- **Default training path may not run as released.** As read, IR-WM's
  semi-coupled config reads a prediction key that its detector does not return,
  so training raises a KeyError. The paper does not mention this. The fully
  decoupled config uses a class that does return the key ([IR-WM card, Paper vs
  code](/papers/2510.16729)). Check it before reproducing the default result.
- **Imitation term form differs in code.** The paper writes the imitation term
  as a squared L2 norm [arxiv:2510.16729 eq. 13]. The released code takes an
  unsquared Euclidean distance ([IR-WM card, Paper vs code](/papers/2510.16729)).
  The two give different gradients, so pick one and state it.
- **Released code differs from the paper.** Octo's paper gives a pretraining
  batch of 2048, and the released config uses 512 [arxiv:2405.12213 §3.4]. The
  code also clips diffusion targets and sampler iterates to [-5, 5] and scales
  the loss by the action dimension. The paper says nothing about either
  [arxiv:2405.12213 §3.3] ([Octo card, Paper vs code](/papers/2405.12213)).
- **Staged training needs data the release does not have.** DexVLA's repository
  has no stage-1 training code, and its sub-step labels (Gemini 2.0
  segmentation plus hand-written sub-step lists) are not released
  [arxiv:2502.05855 app. C.2]. The stage-1 learning rate and data pipeline
  cannot be checked against code ([DexVLA card, Paper vs code](/papers/2502.05855)).
- **Expert labels depend on the states the solver visits.** In learn2branch the
  labels come from strong branching at nodes of a tree that the paper's code
  grows with pseudo-cost branching, and the expert is queried at a sample of
  nodes [arxiv:1906.01629 §4.1]. The learned policy's inputs depend on that
  choice ([learn2branch card, Paper vs code](/papers/1906.01629)). The paper
  re-implements strong branching in SCIP so that it has no side effects
  [arxiv:1906.01629 suppl. §1].
- **Node counts disagree within one paper.** Learn2branch's GCNN node counts for
  hard set covering are 29,981 in one table and 56,348 in another, while the
  times agree to rounding [arxiv:1906.01629 tab. 2] [arxiv:1906.01629 tab. 3].
  Cite the times, and treat the node counts as uncertain.
- **Seed variance.** AnyTeleop's imitation results use three seeds, and several
  of its gaps are smaller than the standard deviations [arxiv:2307.04577 tab. 3].
- **Paper and code disagree in places.** Examples: Perceiver-Actor's batch size
  (16 in the paper, 8 by default in the code), OpenVLA's data-mixture weights
  (the paper's shares differ from the released weights), and WebShop's test-time
  search choice (a random pick from the top 5 in the paper, top 1 in the code).
  The cards list each difference under Paper vs code
  ([Perceiver-Actor](/papers/2209.05451), [OpenVLA](/papers/2406.09246),
  [WebShop](/papers/2207.01206)).
- **Training settings differ from the paper.** Touch in the Wild's paper says
  the masking applies to 95% of samples independently, but the code draws it
  once per batch. The paper gives 60 training epochs for the policy, but the
  default config runs 200. In Stage 2 the code uses a CLIP learning rate of
  2×10⁻⁵ where the paper says 3×10⁻⁵ [arxiv:2507.15062 §4.3] [arxiv:2507.15062
  §5.2]. Choose one reading and report it ([Touch in the Wild card, Paper vs
  code](/papers/2507.15062)).
- **Released configs differ from the paper's runs.** BC-Z's paper gives a
  multi-task learning rate of 5e-3 and batch 4096 on a TPU pod, while its
  released config uses 2.5e-4 and batch 32 [arxiv:2202.02005 app. D]. DROID's
  default config samples with DDPM, while the paper and its generator use
  DDIM, and the generator leaves the 50/50 mix to the user because the
  two-dataset weights are commented out [arxiv:2403.12945 app. F.1]. The
  card's Paper vs code sections give the file-level evidence ([BC-Z](/papers/2202.02005),
  [DROID](/papers/2403.12945)).
- **Held-out count.** BC-Z's text says 29 held-out tasks, while its table has 28
  rows, and the card uses 28 as the count the table supports [arxiv:2202.02005
  tab. 2]. Name the count when you quote its 32% figure.
- **Imitation weight depends on reading.** TD3+BC's λ averages |Q| over dataset
  actions in the paper, but over the policy's actions in the released code
  [arxiv:2106.06860 eq. 5]. The code also averages the imitation error over
  action dimensions, while the paper does not say whether it sums them
  ([TD3+BC card, Paper vs code](/papers/2106.06860)). A sum-based implementation
  puts more weight on imitation.
- **Released token layout differs from the paper.** The paper's RT-1 variant
  uses 8 action dimensions [arxiv:2310.08864 §4.2], while the released code
  emits 11 tokens. RT-2-X code and checkpoints are not in that repository
  ([Open X-Embodiment card, Paper vs code](/papers/2310.08864)).
- **Planner-dependent imitation term.** REAP's anchor reward is defined by states
  that have a feasible planner path, and the conclusion says planner queries are
  still needed in training [arxiv:2605.08713 §4.3] [arxiv:2605.08713 §6]. An
  implementation without a planner has no anchor reward. The paper also does not
  say whether the planner probability is decided per episode, per step or per
  batch [arxiv:2605.08713 §5.3].
- **Training objective versus code.** DiT4DiT's paper trains both diffusion
  transformers from one loss, but the released code stops the action gradient at
  the video features, so the video model learns only from its own loss. The paper
  also gives no value for the feature timestep or the loss weight
  [arxiv:2603.10448 §4.3] ([DiT4DiT card, Paper vs code](/papers/2603.10448)).
  Choose one reading and report it.
- **Two inference recipes.** Multi-Game DT's main method (κ = 10, top-85 cut-off)
  differs from the recipe in its released code (the best of 128 return samples,
  top-50 cut-off) [arxiv:2205.15241 §3.4] [arxiv:2205.15241 app. B.3]. The
  headline 126% does not name the recipe ([Multi-Game DT card, Pitfalls](/papers/2205.15241)).
- **Lookup scale is unspecified.** VINN's softmin weights are exp(−distance)
  with no temperature, so on raw 2048-d embeddings the average can collapse onto
  one neighbour. The paper does not say whether embeddings are normalised
  [arxiv:2112.01511 §3.2]. Its robot results are 30 rollouts with no confidence
  intervals [arxiv:2112.01511 tab. I].
- **Partial credit in Voltron.** Real-robot scores add 0.25 for each milestone
  reached, so they are not success rates [arxiv:2302.12766 §5.4]. Its control
  results are low in absolute terms [arxiv:2302.12766 §5.3].
- **Tiny test sets for the LLM agent.** On 50 WebShop tasks one success is 0.02.
  The gap between GPT-4 with one opinion and with five (0.300 against 0.320) is
  one task, and the rows are single runs at low temperature [arxiv:2306.02224 §3.1.1].
  The paper's ALFWorld text and table also swap the Claude and GPT-3.5 figures, so
  the table values are used here [arxiv:2306.02224 §3.3.2] [arxiv:2306.02224 tab. 2].
  The expert is also trained on the same environments it advises on
  [arxiv:2306.02224 §2.1.1] [arxiv:2306.02224 §2.1.2].
- **Driving headline depends on a configuration.** PGP's main-table row matches an
  encoder without GNN layers, while its decoder ablation uses two GAT layers. The
  card asks the reader to pick one setting and state it [arxiv:2106.15004 tab. 1]
  [arxiv:2106.15004 tab. 3].
- **Licences of the code.** The Hiking code is CC BY-NC 4.0, so it is
  non-commercial ([Hiking card](/papers/2601.07718)). The DexVLA code carries a
  CC BY-NC style licence, so it is non-commercial too, and the card cites it for
  facts only ([DexVLA card, Pitfalls](/papers/2502.05855)). The pedagogical IRL
  repository has no licence file ([Showing versus Doing card](/papers/2610.14004)).
  Check the licence of each repository before reuse.
- **Loss not named in OTTER.** The paper does not name a training loss
  [arxiv:2503.03734 §3.2]. Its released code uses L1 unless a flag is set, but
  the config it reads has no such field, so the model does not construct as
  cloned ([OTTER card, Paper vs code](/papers/2503.03734)). Treat the loss as
  open until it is checked.
- **Partial scores are not success rates.** The OXE-AugE physical scores
  average 0.3, 0.5, 0.8 and 1.0 per trial [arxiv:2512.13100 app. A.5], and its
  24% and 45% gains have no stated unit or denominator [arxiv:2512.13100 §6.2].
  OTTER's pick-and-place trials also give partial credit, 0.5 for the pick and
  1 for the place [arxiv:2503.03734 app. A.2]. Compare these only with other
  partial-credit results.
- **Augmentation assumptions.** OXE-AugE assumes similar grippers and that
  every arm can do the task with the same strategy [arxiv:2512.13100 §3]
  [arxiv:2512.13100 §7]. Only "over 95%" of trajectories reach under 0.25 cm
  replay error for most datasets, and the rest are kept without a stated bound
  [arxiv:2512.13100 §4.2 (2)] fn. 1.
- **Corrections are not a controlled comparison.** HACTS attributes its gain in
  the failure-correction setting to correction data covering failed states, a
  claim the paper does not test directly [arxiv:2503.24070 §4.3]. Its "without
  any performance drop" statement rests on single 10-rollout comparisons
  [arxiv:2503.24070 §4.3].

- **Simulation-only evidence.** Several headline results are simulated. PDP's
  results are all in simulation [arxiv:2406.00960 §5]. The RP1M benchmark is
  simulated [arxiv:2408.11048 §6], as are the humanoid study [arxiv:2411.01349
  §5], LiveNet's comparisons [arxiv:2412.04659 §5.2] and FeudalNav
  [arxiv:2402.12498 §8]. AMR's forklift result is simulated only
  [arxiv:2411.14770 §V-D].
- **Discriminator sign in StepAgent.** As written, its objective pushes the
  discriminator toward high scores on the agent's own pairs, and the reward is
  the log of that score. This is the reverse of the usual GAIL convention. The
  paper also uses D in three different ways [arxiv:2411.03817 eq. 10]
  [arxiv:2411.03817 alg. 1] [arxiv:2411.03817 §5]. Check the sign and the reward
  form before implementing the inverse-RL variant.
- **Noise details are missing.** PDP and LiCS add noise to executed actions and
  train on clean targets. PDP leaves the noise level for tracking and
  text-to-motion unstated [arxiv:2406.00960 tab. 1], and LiCS does not give the
  unit of its σ = 0.25 [arxiv:2406.14947 §IV-A]. Record the choice you make.
- **Inference cost of diffusion heads.** PDP's inference runs K times the cost
  of an MLP [arxiv:2406.00960 §6]. HDP runs 100 Guider and 10 Actor denoising
  steps per control step [arxiv:2411.12982 tab. 9]. ViDEN's control step also
  runs several U-Net passes [arxiv:2412.20226 §II-D].
- **Small real-robot samples.** RcP's real experiment has one attempt per network
  and condition at each of 50 target locations [arxiv:2411.04331 §4.3]. TaKSIE's
  real-robot table uses 5 trials per task [arxiv:2410.11013 tab. 3]. SAME's
  routing and expert ablations are single runs with no seeds [arxiv:2412.05552
  tab. 2]. ViDEN's 25 trials per level all come from one robot
  [arxiv:2412.20226 §III-C].
- **Baselines changed or unstated.** HDP's Diffusion Policy baseline is smaller
  than the published one [arxiv:2411.12982 §6.2]. The LLM-guided driving study
  uses hosted models whose versions are not stated, so its numbers may not
  repeat [arxiv:2411.14256 §3.2].
- **Survey classes are not results.** The offline RL survey's imitation class
  and its heat map collect scores from different papers, protocols and
  environments, and take class maxima. They show coverage, not a controlled
  ranking of imitation against other methods [arxiv:2203.01387 §VI-C]
  [arxiv:2203.01387 fig. 5]. Take the method's own card for its numbers.

## Papers

- [Exact Combinatorial Optimization with Graph Convolutional Neural Networks](/papers/1906.01629): a graph network that imitates strong branching to choose variables in a mixed-integer solver.
- [One Thousand and One Hours: Self-driving Motion Prediction Dataset](/papers/2006.14480): an open driving dataset whose planning baseline imitates logged ego trajectories with perturbations.
- [WebShop: Towards Scalable Real-World Web Interaction with Grounded Language Agents](/papers/2207.01206): a shopping environment whose agents are first trained by imitation, then tuned with policy gradient.
- [Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation](/papers/2209.05451): a voxel keyframe behaviour-cloning agent for multi-task manipulation.
- [VIMA: General Robot Manipulation with Multimodal Prompts](/papers/2210.03094): a prompt-conditioned behaviour-cloning policy trained on scripted expert data.
- [Orbit: A Unified Simulation Framework for Interactive Robot Learning Environments](/papers/2301.04195): a simulator that collects teleoperated demonstrations and reports behaviour-cloning results on a lift-cube task.
- [Learning Universal Policies via Text-Guided Video Generation](/papers/2302.00111): a planner that generates video from text and reads actions from the frames.
- [LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning](/papers/2306.03310): a benchmark of 130 language-conditioned tasks with demonstrations, and a comparison of lifelong learning methods.
- [AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System](/papers/2307.04577): a vision-based teleoperation stack for collecting demonstrations across robots.
- [OpenVLA: An Open-Source Vision-Language-Action Model](/papers/2406.09246): a 7B open vision-language-action model trained on robot demonstrations, with a LoRA fine-tuning recipe.
- [Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion](/papers/2407.01392): a sequence model with per-token noise levels, used for planning and robot imitation.
- [FAST: Efficient Action Tokenization for Vision-Language-Action Models](/papers/2501.09747): a compressed action tokenizer that makes autoregressive VLA training work on high-frequency dexterous tasks.
- [Hiking in the Wild: A Scalable Perceptive Parkour Framework for Humanoids](/papers/2601.07718): a humanoid parkour policy trained with RL and an adversarial style reward.
- [EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies](/papers/2606.20092): a VLA policy with a learned keyframe memory for briefly visible cues.
- [Showing versus Doing: Teaching by Demonstration](/papers/2610.14004): a model of teaching by demonstration, with two human studies and a reward-recovery comparison.
- [The Surprising Effectiveness of Representation Learning for Visual Imitation](/papers/2112.01511): a nearest-neighbour lookup over demonstration frames, on a BYOL-tuned encoder.
- [Multi-Game Decision Transformers](/papers/2205.15241): one return-conditioned transformer trained offline on 41 Atari games.
- [Language-Driven Representation Learning for Robotics](/papers/2302.12766): Voltron, a video-caption masked autoencoder whose frozen features feed robot imitation heads.
- [DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control](/papers/2603.10448): a video diffusion transformer coupled to an action head, trained with flow matching.
- [Touch in the Wild: Learning Fine-Grained Manipulation with a Portable Visuo-Tactile Gripper](/papers/2507.15062): a handheld tactile gripper, a visuo-tactile encoder pretrained by masked reconstruction, and a diffusion policy trained by behaviour cloning.
- [Multimodal Trajectory Prediction Conditioned on Lane-Graph Traversals](/papers/2106.15004): PGP, a route policy trained by behaviour cloning on lane graphs, then trajectory decoding for driving.
- [Auto-GPT for Online Decision Making: Benchmarks and Additional Opinions](/papers/2306.02224): LLM agents on WebShop and ALFWorld that receive a trained imitator's suggested action in the prompt.
- [REAP: Reinforcement-Learning End-to-End Autonomous Parking with Gaussian Splatting Simulator for Real2Sim2Real Transfer](/papers/2605.08713): a camera-based SAC parking policy in a Gaussian-splatting simulator, with planner behaviour cloning mixed into the actor loss.
- [A Minimalist Approach to Offline Reinforcement Learning](/papers/2106.06860): TD3+BC, a TD3 actor with a behaviour-cloning term for learning from fixed logs.
- [Open X-Embodiment: Robotic Learning Datasets and RT-X Models](/papers/2310.08864): pooled robot datasets from 22 embodiments, and RT-1-X and RT-2-X policies trained on them.
- [Octo: An Open-Source Generalist Robot Policy](/papers/2405.12213): a pretrained transformer policy with a diffusion action head, finetuned to new sensors and action spaces.
- [DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control](/papers/2502.05855): a VLM with a diffusion action expert, trained in three stages with sub-step text.
- [Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success](/papers/2502.19645): OpenVLA-OFT, a fine-tuning recipe with parallel decoding, action chunks and an L1 head.
- [GR00T N1: An Open Foundation Model for Generalist Humanoid Robots](/papers/2503.14734): a humanoid VLA with a flow-matching diffusion transformer, pretrained on real, simulated and video-derived data.
- [Vision-Centric 4D Occupancy Forecasting and Planning via Implicit Residual World Models](/papers/2510.16729): a driving world model that predicts residual BEV changes and plans with an imitation-plus-collision loss.
- [BC-Z: Zero-Shot Task Generalization with Robotic Imitation Learning](/papers/2202.02005): a language- or video-conditioned behaviour-cloning policy trained on 100 real-robot tasks.
- [DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset](/papers/2403.12945): a shared-hardware manipulation dataset, and a diffusion policy co-trained with it.
- [OTTER: A Vision-Language-Action Model with Text-Aware Visual Feature Extraction](/papers/2503.03734): a VLA that keeps CLIP frozen and selects image tokens by similarity to the instruction.
- [HACTS: a Human-As-Copilot Teleoperation System for Robot Learning](/papers/2503.24070): a bilateral leader arm whose recorded human corrections feed imitation and RL fine-tuning.
- [OXE-AugE: A Large-Scale Robot Augmentation of OXE for Scaling Cross-Embodiment Policy Learning](/papers/2512.13100): cross-painting of Open X-Embodiment demonstrations onto other arms, used to fine-tune OpenVLA-OFT and pi0.
- [Feudal Networks for Visual Navigation](/papers/2402.12498): a three-level navigator trained on human point-clicks, with no RL, odometry or metric map.
- [Autonomous Driving Small-Scale Cars: A Survey of Recent Development](/papers/2404.06229): a survey of small-scale autonomous-car platforms and driving tasks, with no experiments.
- [PDP: Physics-Based Character Animation via Diffusion Policy](/papers/2406.00960): a diffusion policy for simulated characters, cloned from noisy rollouts of RL experts.
- [LiCS: Navigation using Learned-imitation on Cluttered Space](/papers/2406.14947): a small Transformer that imitates a local planner for wheeled navigation, with noise-augmented data.
- [ManiWAV: Learning Robot Manipulation from In-the-Wild Audio-Visual Data](/papers/2406.19464): a diffusion policy that fuses camera and contact-audio features, with data collected by a handheld gripper.
- [RP1M: A Large-Scale Motion Dataset for Piano Playing with Bi-Manual Dexterous Robot Hands](/papers/2408.11048): a simulated piano dataset from RL specialists, with multi-task imitation baselines.
- [Scaling Diffusion Policy in Transformer to 1 Billion Parameters for Robotic Manipulation](/papers/2409.14411): ScaleDP, a transformer diffusion policy with adaptive layer norm, scaled from 10M to 1B parameters.
- [Incorporating Task Progress Knowledge for Subgoal Generation in Robotic Manipulation through Image Edits](/papers/2410.11013): TaKSIE, progress-selected visual subgoals with a subgoal-conditioned policy.
- [The Role of Domain Randomization in Training Diffusion Policies for Whole-Body Humanoid Control](/papers/2411.01349): diffusion policies trained on humanoid walking data from an AMP expert, varying randomisation and dataset size.
- [From Novice to Expert: LLM Agent Policy Optimization via Step-wise Reinforcement Learning](/papers/2411.03817): StepAgent, a step-wise reward from expert and agent actions for LLM agents.
- [Object-Centric Dexterous Manipulation from Human Motion Data](/papers/2411.04005): a wrist planner trained by behaviour cloning on human motion, with a PPO finger controller.
- [Raising Body Ownership in End-to-End Visuomotor Policy Learning via Robot-Centric Pooling](/papers/2411.04331): RcP, a pooling layer that selects the image features matching the robot's state.
- [Moving Forward: A Review of Autonomous Driving Software and Hardware Systems](/papers/2411.10291): a survey of driving software and hardware, with a throughput benchmark of three end-to-end models.
- [Hierarchical Diffusion Policy: manipulation trajectory generation via contact guidance](/papers/2411.12982): HDP, a two-level diffusion policy with contact guidance and a Q-learning term.
- [Generalizing End-To-End Autonomous Driving In Real-World Environments Using Zero-Shot LLMs](/papers/2411.14256): an imitation driving policy with a frozen multimodal LLM that gives direction hints.
- [Aim My Robot: Precision Local Navigation to Any Object](/papers/2411.14770): AMR, a transformer that drives to a centimetre-precise pose relative to an object.
- [LiveNet](/papers/2412.04659): a neural controller that imitates MPC for two robots, with a QP safety layer.
- [SAME: Learning Generic Language-Guided Visual Navigation with State-Adaptive Mixture of Experts](/papers/2412.05552): a mixture of experts in DUET's cross-attention, trained by DAgger on seven VLN tasks.
- [RLDG: Robotic Generalist Policy Distillation via Reinforcement Learning](/papers/2412.09858): distils successful RL-expert rollouts into OpenVLA and Octo.
- [Emma-X: An Embodied Multimodal Action Model with Grounded Chain of Thought and Look-ahead Spatial Reasoning](/papers/2412.11974): OpenVLA fine-tuned on demonstrations relabelled with subtasks, reasoning and movement plans.
- [Embodiment-Agnostic Navigation Policy Trained with Visual Demonstrations](/papers/2412.20226): ViDEN, a diffusion navigation policy trained on hand-held depth-camera demonstrations.
- [HG-DAgger: Interactive Imitation Learning with Human Experts](/papers/1810.02890): a DAgger variant where the human decides when to take over, with an ensemble doubt threshold.
- [Trajectory-guided Control Prediction for End-to-end Autonomous Driving: A Simple yet Strong Baseline](/papers/2206.08129): a driving network with waypoint and control branches, with attention guided by the waypoints.
- [NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration](/papers/2310.07896): one diffusion policy that explores when the goal is masked and reaches a goal image when it is shown.
- [Imitating and Finetuning Model Predictive Control for Robust and Symmetric Quadrupedal Locomotion](/papers/2311.02304): a DAgger clone of a DDP-based MPC, fine-tuned with PPO on rough terrain.
- [A Survey on Offline Reinforcement Learning: Taxonomy, Review, and Open Problems](/papers/2203.01387): a survey that classes offline RL methods, imitation among them, and reviews benchmarks.
- [Safety-Enhanced Autonomous Driving Using Interpretable Sensor Fusion Transformer](/papers/2207.14024): InterFuser, a multi-sensor driving transformer that imitates a privileged expert, with a rule-based speed safety controller.
- [TrafficSim: Learning to Simulate Realistic Multi-Agent Behaviors](/papers/2101.06557): a joint traffic policy trained from driving logs by unrolling it in closed loop, with an imitation term and a collision penalty.
- [A Survey of Imitation Learning: Algorithms, Recent Developments, and Challenges](/papers/2309.02473): a survey of behavioural cloning, IRL, adversarial IL and imitation from observation, with a table of 66 works.
- [FILIC: Dual-Loop Force-Guided Imitation Learning with Impedance Torque Control for Contact-Rich Manipulation Tasks](/papers/2509.17053): a transformer policy with an estimated end-effector wrench and a 2 kHz impedance loop, with no force sensor.