A Survey of Imitation Learning: Algorithms, Recent Developments, and Challenges

In plain words

Writing down every rule a robot or program should follow is slow, and so is designing a reward score that makes it behave well. Imitation learning takes a different route: show the system what a skilled person does, and let it learn to do the same.

The simplest version copies the demonstrations as ordinary supervised learning. The trouble is that a small early mistake leads the copy into situations the expert never showed, and it may not know how to get back. The survey walks through the main fixes. One asks the expert for corrections during training. Another infers the goal behind the demonstrations. A third trains a judge that rewards behaviour that looks like the expert's. A fourth learns from videos that do not record the actions. The survey also covers demonstrations that are noisy or come from a different body, viewpoint or setup. It does not run experiments; it organises the field and points out open problems.

TL;DR

• A 2023 survey of imitation learning (IL) for robotics and AI §I.
• It groups methods into behavioural cloning (BC), inverse reinforcement learning (IRL), adversarial IL (AIL) and imitation from observation (IfO), then covers challenges §I §II §III §IV §V §VI.
• The central technical issue for BC is covariate shift: the learner drifts into states the expert never demonstrated §II.
• One table lists 66 works with learning type, online or offline setting, online-expert use and application tab. I.
• The survey reports no experiments. Several places where its text and table disagree are listed under Pitfalls.

Problem

Hand-coding behaviour or reward functions becomes impractical in unstructured settings such as driving, aerial robotics and language tasks §I. Imitation learning replaces that effort with demonstrations §I.

The survey singles out these difficulties:

• Covariate shift in BC. The learner is trained on states visited by the expert but tested on states visited under its own actions. Errors then compound. This matters most in safety-critical settings such as driving §II.
• Cost and safety of IRL. IRL needs repeated interaction with the environment to estimate the reward, which can be expensive and unsafe in high-risk applications §III.
• Ambiguity in IRL. A policy can be optimal for many reward functions, so IRL methods must add structure to choose one §III.
• Training instability in AIL. Adversarial, min-max training can suffer from vanishing gradients and failure to converge §IV.
• Missing actions in IfO. Videos and similar sources give states but not actions, which limits direct copying §V.
• Assumptions in practice. Many methods assume the demonstrations are optimal §VI-A, that expert and agent share state and action spaces §VI-B, or, for IRL, that the transition model is known §III.

Method

The survey is a taxonomy, not a method. It covers BC, then IRL, AIL and IfO, then challenges §I. The categories below summarise what it says about each family. Sections use the IEEE numbering the compiled paper prints (I to VII; see Pitfalls).

1. Behavioural cloning and covariate shift §II fig. 2 fig. 3 fig. 4

BC treats imitation as supervised learning from expert state-action pairs. It needs no model of the dynamics §II. The survey groups fixes for covariate shift into three areas fig. 2:

• Interactive IL assumes an online expert. DAgger relabels the agent's own visited states with expert actions §II. Human-gated methods let the expert decide when to intervene (HG-DAgger). Robot-gated methods let the robot ask for help (SafeDAgger, LazyDAgger, ThriftyDAgger). ThriftyDAgger caps the number of interventions with a user-set budget §II.
• Support-keeping RL with a fixed reward needs no online expert. One method scores how close a state-action pair is to the demonstrated support using kernel PCA §II. SQIL gives +1 for transitions inside demonstrations and 0 otherwise §II. A disagreement method uses the variance of an ensemble as a cost, combined with a supervised BC loss §II.
• Constrained IL keeps the agent in known regions without an expert or RL. One driving method adds extra losses and synthetic perturbations of expert trajectories §II. A learned error detector resets or stops the agent in unfamiliar states §II. An offline method penalises regions of the model that the data does not cover §II.

Causal confusion (copycat) is treated separately. The learner copies previous expert actions instead of finding their causes §II fig. 3. One line of work learns an adversarial representation that drops the previous-action information; a later extension applies this to image observations with a memory module §II.

An implicit BC model treats the policy as an energy function over observation-action pairs and picks the lowest-energy action. The survey notes that it is slower than explicit BC but can represent sharp discontinuities §II.

2. Inverse reinforcement learning §III fig. 4

IRL infers a reward from demonstrations, then trains a policy with RL §III. The survey sorts IRL methods by how they resolve reward ambiguity:

• Maximum margin. The reward should explain the demonstrated policy better than others by a margin. Methods include a linear-programming formulation §III, maximum-margin planning, its functional-gradient extensions §III, feature-expectation margins §III and a game-theoretic variant that can beat the expert §III.
• Maximum entropy. The method maximises the entropy of the resulting policy, using a distribution over trajectories §III. The survey covers the original method, its continuous-space path-integral extensions, a deep version that learns features with neural networks, and guided cost learning, which learns the reward in the inner loop and uses raw states §III.
• Bayesian. A prior over rewards is updated by expert actions. Variants use MCMC for the posterior mean, MAP inference, a Gaussian-process reward, a preference-based likelihood that avoids solving an MDP per sample §III fig. 4, and approximate variational inference §III.

3. Adversarial imitation learning §IV

A discriminator learns to tell expert from agent trajectories. The agent is rewarded for fooling it, so the reward is derived from the discriminator's confusion §IV. The survey notes that this avoids solving a full RL problem in every iteration §IV. Later work changes the discriminator loss §IV, switches to off-policy learning §IV, or replaces f-divergences with a Wasserstein distance §IV. A primal-dual approximation of the Wasserstein distance is another option §IV.

4. Imitation from observation §V fig. 5 fig. 6 fig. 7 fig. 8

IfO learns from states without actions. The survey lists these methods:

• Context translation maps demonstrations from a third-person view to the robot's view. A reward penalises deviating from the translated features §V fig. 5. It needs time-aligned demonstrations and many of them §V.
• Time-contrastive networks learn a viewpoint-invariant frame embedding with a triplet loss; the reward is the Euclidean distance in that embedding §V fig. 6. Multi-viewpoint video is required §V.
• Behavioural cloning from observation learns an inverse dynamics model by letting a random-policy agent interact with the environment, labels the demonstrations with inferred actions, then runs BC §V. The interaction cost is high in high-dimensional problems §V.
• Latent-action methods learn a forward dynamics model offline and use a few environment interactions to map latent actions to real ones §V.
• GAIfO applies the GAIL objective to state transitions. It handles demonstrations that are not time-aligned, but struggles when dynamics differ §V.
• Conditional transition matching is non-adversarial and matches conditional transition probabilities §V.
• Cross-domain state maps align unpaired demonstrations with adversarial training and a learned temporal position, then apply BCO §V.
• YouTube imitation learns a self-supervised vision and audio embedding, places checkpoints along one video, and rewards the agent for reaching them §V fig. 7.
• Ranked-demonstration reward learns a state-based reward so that higher- ranked trajectories get more return §V.
• Driving from web video trains a bird's-eye-view model on a small labelled set, pseudo-labels a large unlabelled set, and fine-tunes the policy §V.

5. Challenges §VI fig. 8 fig. 9 tab. I

Imperfect demonstrations §VI-A. Two-step importance weighting (2IWIL) and an imperfect-demonstration GAIL (IC-GAIL) use confidence scores §VI-A. A reweighted BC scheme uses the learned policy as sample weights; the survey notes it converges only when optimal demonstrations are the majority §VI-A. A weighted GAIL estimates each demonstration's weight from the discriminator and the policy §VI-A. A regularised distribution-matching method mixes expert and imperfect data §VI-A. ILEED models each demonstrator's expertise per state with the inner product of state and demonstrator embeddings §VI-A fig. 8.

Domain discrepancies §VI-B fig. 9. The survey separates dynamics, viewpoint and embodiment mismatch §VI-B. Robust IL for perturbed dynamics minimises a weighted Jensen-Shannon objective over several expert policies §VI-B fig. 9. An adversarial viewpoint method removes viewpoint information without proxy tasks §VI-B. Cross-embodiment IRL uses temporal cycle-consistency embeddings and a goal-distance reward §VI-B. Another method uses Gromov-Wasserstein distance between occupancy measures to compute pseudo-rewards §VI-B. Unpaired state and action maps allow zero-shot imitation with an online expert §VI-B.

Interactive DAgger loop

The survey describes DAgger in prose §II. The pseudocode below follows that description. The number of rounds is not stated.

# Dataset aggregation (DAgger)  [§II]
# Input:  expert (online, relabels on request); demonstrations D
# Output: policy trained on the aggregated data

def dagger(expert, demonstrations, rounds):
    # Start from the expert's demonstrations  [§II]
    dataset = demonstrations
    policy = train_bc(dataset)
    for round in range(rounds):
        # Run the current policy and record the states it visits  [§II]
        states = run_policy(policy)
        # The expert gives the action it would take in each visited state  [§II]
        labels = expert.relabel(states)
        # Aggregate the new pairs with all earlier data  [§II]
        dataset = dataset + (states, labels)
        policy = train_bc(dataset)
    return policy

Taxonomy at a glance:

Hyperparameters

Not applicable. The survey proposes and runs no method. The sections that describe cited methods give no hyperparameter values §II §III §IV §V §VI-A §VI-B.

Results

Not applicable. The survey reports no experiments of its own. Its only quantitative content is tab. I, which lists per work the datasets (simulated or real), input types (state or image), online or offline setting, online-expert use and application domain. Performance statements in the text are attributed to the cited works. Examples: SQIL is said to beat simple BC with few demonstrations §II, and GAIL-family methods are said to show statistically significant gains on several benchmarks §IV. The survey gives no numbers for either claim.

Paper vs code

No official code. The abstract, the paper body and the arXiv metadata contain no repository link. The LaTeX source has no URL except the url package import. I did not run a GitHub search: a survey has no implementation, and repositories that collect surveyed papers are curated lists rather than code.

Pitfalls

• Table and text disagree on online-expert use. tab. I marks the kim2020domain entry (Kim et al.) as having no online expert. The text says the framework works "while having access to an online expert" §VI-B. Check the original before relying on either.
• Table and text disagree on method families. tab. I puts the disagreement method, the ChauffeurNet-style driving method and the kernel-PCA support method under "IL via Expert Support Estimation". The text describes the first as an ensemble-variance cost §II and the second as extra losses with synthetic perturbations §II.
• Novelty claim conflicts with the text. IfO is called "novel" in the introduction §I and in the conclusion §VII. Section V says state-only demonstrations are "not new" §V.
• Probable typo. GAIL's divergence is written "Shannon-Jensen" §IV. The usual name is Jensen-Shannon. Confirm against the original method before using the term.
• Possibly misplaced citation. The description of 2IWIL cites a DemoDICE- style work for its reweighted GAIL step §VI-A. Check which work the step comes from.
• Two tables carry number I. The survey table is split into two parts, both captioned "Table I" (the second is the continuation). Cite as tab. I and say which part.
• Section numbers are inferred. The LaTeX source has no section numbers. This card uses the IEEE Roman numerals the compiled paper prints: I to VII, with VI-A and VI-B for the two challenge subsections. The paper's own cross reference to "section III and IV" for IRL and AIL confirms the scheme.
• Hidden costs. BCO needs a lot of random-policy interaction to learn the inverse dynamics model §V. Reweighted BC converges only when optimal demonstrations are the majority §VI-A. IRL requires repeated environment interaction §III.

Lineage

• Surveys positioned against (no cards; named only): Schaal (1999) on IL for humanoid robots; Osa et al. (2018), an algorithmic view; Hussein et al. (2017), the design options of the IL process; Le Mero et al. (2022), IL for end-to-end driving §I.
• Key methods the survey reviews, which later cards may link: DAgger §II, maximum-entropy IRL §III, GAIL §IV and AIRL §IV.
• Compared against: the 66 works in tab. I.
• Superseded by: not stated in the paper, and no later version is in this wiki's sources.