---
kind: concept
slug: long-short-term-memory
title: Long short-term memory (LSTM)
summary: >-
  A gated recurrent cell with an additive cell-state update. xLSTM adds exponential gating and matrix memory, and reports lower SlimPajama perplexity than Llama and Mamba at 15B tokens. LSTMs also classify skeletons and tactile series, decode dialogue, correct vehicle forces, decode EMG, detect anomalies and run PPO locomotion policies. Results are author-reported.
tags: [recurrent-networks, sequence-modelling, language-modeling, exponential-gating, reinforcement-learning, locomotion, tactile-sensing, time-series, teleoperation, vehicle-dynamics, emg, action-recognition, dialogue, anomaly-detection, speech-translation]
papers: ["2405.04517", "1704.02581", "2204.04340", "2204.06343", "2402.10220", "2412.00581", "2412.17991", "1705.05414", "2107.12626", "1612.01744"]
related: [recurrent-neural-network, state-space-model, linear-attention, attention-mechanism, backpropagation-through-time, proximal-policy-optimization, sim-to-real-transfer, support-vector-machine, time-series-classification, time-series-forecasting, temporal-convolutional-network, convolutional-recurrent-network, residual-learning, dialogue-generation, knowledge-base-question-answering, anomaly-detection, autoencoder, neural-machine-translation]
status: draft
updated_at: 2026-11-01
---

## Overview

An LSTM keeps a cell state that is updated additively and scaled by gates.
That additive path is the "constant error carousel" the paper builds on
[arxiv:2405.04517 §2.1]. The paper names three limits of the classic LSTM:
it cannot revise a stored value once a better match appears later in the
sequence, it compresses everything into scalar cell states, and its
hidden-to-hidden connections force sequential processing
[arxiv:2405.04517 §1]. The variants below address these limits. The
invariant they all keep is the gated, additive cell-state update.

LSTMs are also used as a recurrent controller rather than a language model.
The bipedal-locomotion card uses a two-layer, 128-unit LSTM as the policy
that maps proprioceptive input and a clock to joint set points
[arxiv:2204.04340 §III-A]. That use is covered in its own variant below.

Outside language modelling and locomotion, the LSTM is a classifier, a
forecaster, a residual corrector and a regressor over sensor streams. It is
also the sequence predictor inside an unsupervised anomaly detector. It is
the encoder and decoder of a task-oriented dialogue model and of a
speech-to-text translator, and the per-step reader of skeleton joints. Each of these has its own variant below,
and each card reports results on its own task, so the numbers are not
comparable across variants.

## Variants

### Exponential gating with normalizer and stabilizer

The input gate, and optionally the forget gate, use the exponential
function. A normalizer state tracks the overall gate strength, and a
stabilizer state is subtracted inside the exponentials so they do not
overflow. The stabilized and unstabilized forms give the same outputs and
parameter gradients [arxiv:2405.04517 §2.2; app. A.2]. In the 15B ablation,
adding exponential gating to the vanilla-style backbone lowers perplexity
from 26.01 to 17.70, and adding matrix memory lowers it further to 13.48
[arxiv:2405.04517 tab. 2].

### sLSTM: scalar memory with memory mixing

The sLSTM keeps the scalar cell and adds recurrent matrices from the previous
hidden state to every gate. These matrices are block-diagonal, one block per
head, so heads do not mix [arxiv:2405.04517 §2.2; app. A.2]. Because of this
recurrence it is not parallelisable over the sequence
[arxiv:2405.04517 §5 (i)]. Card: [xLSTM: Extended Long Short-Term Memory](/papers/2405.04517).

### mLSTM: matrix memory with covariance update

The mLSTM replaces the scalar cell with a d x d matrix. Keys and values are
written with an outer-product (covariance) rule and read back with a query.
Its gates depend only on the input, not on the previous output, so training
can use a parallel form over the whole sequence [arxiv:2405.04517 §2.3;
app. A.3]. Card: [xLSTM: Extended Long Short-Term Memory](/papers/2405.04517).

### xLSTM blocks and the a:b ratio

sLSTM and mLSTM sit in residual blocks with pre-LayerNorm. The notation
xLSTM[a:b] gives the ratio of mLSTM to sLSTM blocks [arxiv:2405.04517 §2.4].
The paper reports two families: xLSTM[1:0], which is mLSTM only, and
xLSTM[7:1], which has sLSTM blocks at a few fixed positions
[arxiv:2405.04517 §4.2; app. B].

### Recurrent policy trained with PPO

The locomotion paper's policy is an LSTM with two hidden layers of 128 units.
Its inputs are pelvis orientation and rotational velocity, joint positions and
velocities, a sine and cosine clock whose period is the cycle time, and the
desired forward velocity. Its outputs are joint set points at 40 Hz
[arxiv:2204.04340 §III-A]. It is trained with PPO using a clipped objective and
gradient clipping [arxiv:2204.04340 §III-A]. Card: [Sim-to-Real Learning for
Bipedal Locomotion Under Unsensed Dynamic Loads](/papers/2204.04340).

The policy input does not change with the load. Only the simulated robot model
changes [arxiv:2204.04340 §III-B]. Training on the specific load raises the
Carry Pole pass rate from 0.0014 for the base policy to 0.9578
[arxiv:2204.04340 tab. 2]. The card records no non-recurrent policy as a
comparison, so these results cannot be credited to the LSTM alone.

### Recurrent classifier on raw squeeze signals

The tactile paper feeds each squeeze's raw sensor series to an LSTM. The
series is zero-padded to a common length. The output at the last non-zero
step goes through two linear layers and a softmax, trained with cross-entropy
and Adam [arxiv:2204.06343 §4.2.3]. The grid covers 2 or 4 layers, 32 to 256
units and a log range of learning rates. The paper trained 336 such models,
and the best were typically 4 layers with 256 units [arxiv:2204.06343 §4.2.3].
A support vector machine on hand-made features is the other main classifier
in the same study [arxiv:2204.06343 §4.2.2]. Card: [Single-grasp Deformable
Object Discrimination](/papers/2204.06343).

### CNN features feeding an LSTM forecaster

The teleoperation paper first classifies six glovebox tasks with a CNN. It
then feeds CNN features into an LSTM that forecasts the next 500 frames, which
is 5 s at 100 Hz [arxiv:2402.10220 §IV-A; §IV-E; tab. III]. The
baseline LSTM instead takes the mean over the 24 input channels at each step
[arxiv:2402.10220 §IV-E.1]. The LSTM stack is 100, 100, 50 and 50 units,
read from a Netron export, with a 200-unit dense layer after it
[arxiv:2402.10220 fig. 8d]. Card: [Towards Semi-Autonomous Robotic Arm
Manipulation Operator Intention Detection](/papers/2402.10220).

### LSTM residual correction to a physics model

The off-road driving paper keeps a bicycle-type physics model for the tyre
and drive forces. An LSTM predicts a correction to those forces from a 0.2 s
history of states, controls and measurements [arxiv:2412.00581 §4; eq. 5].
In the visual variant, a small per-wheel encoder maps terrain features into
extra LSTM inputs, and the whole model is trained end to end
[arxiv:2412.00581 §5; eq. 6; fig. 1]. Card: [Dynamics Modeling using Visual
Terrain Features](/papers/2412.00581).

### Causal LSTM regressor for EMG

The myoelectric paper maps 16 surface-EMG channels, reduced to time-domain
features, to seven wrist and finger angles. Its LSTM has 64 nodes, a sigmoid
output and an MSE loss. The window is 50 ms, and the model trains for 40
epochs with batch size 24 [arxiv:2412.17991 §2.1.6; tab. 1]. The model is
strictly causal, so each prediction uses only earlier EMG
[arxiv:2412.17991 §2.1.6]. After each trial the model is updated on that
trial's data. Both sequential models improve by about 4 degrees RMSE over the
session, which the paper reports as significant [arxiv:2412.17991 §3.3]. Card:
[Online Adaptation for Myographic Control](/papers/2412.17991).

### Attention BiLSTM predictor inside an autoencoder anomaly detector

The CAE-M model trains a convolutional autoencoder jointly with a
bidirectional LSTM. The LSTM reads the earlier time steps of each sample's
feature sequence and predicts the next one. Attention over the step outputs
gives a softmax-weighted sum that serves as the prediction [arxiv:2107.12626
§3.4.1; eq. 13–17]. The LSTM has 512 units, and dropout of 0.2 sits after
the attention. The text merges the two directions by sum, while eq. 13
writes concatenation. The paper does not say whether 512 is per direction
[arxiv:2107.12626 §3.4.1; eq. 13; §4.3]. The LSTM term is weighted 0.5 in the
total loss, next to a linear predictor and the reconstruction term
[arxiv:2107.12626 §3.5; eq. 20].

On three datasets the paper reports the same ablation for attention. Removing
it lowers mF1 from 0.9639 to 0.9213 on PAMAP2, from 0.9961 to 0.9271 on CAP,
and from 0.9960 to 0.9578 on fatigue. These are single values with no spread
[arxiv:2107.12626 tab. 6]. Removing both predictors (λ2 = λ3 = 0) gives 0.8063
on PAMAP2, but that also removes the linear predictor, so it does not isolate
the LSTM [tab. 6]. Against the ten baselines, CAE-M reports the highest mF1
on all three datasets. The LSTM autoencoder baseline reports 0.8296 on
PAMAP2, 0.6671 on CAP and 0.6870 on fatigue [arxiv:2107.12626 tab. 2]. The
paper releases no code and does not state the learning rate, so an
implementation must fill in those values [arxiv:2107.12626 §4.3]. Card:
[Unsupervised Deep Anomaly Detection for Multi-Sensor Time-Series
Signals](/papers/2107.12626).

### Stacked and hierarchical LSTM over skeleton joints

The skeleton paper's temporal stream runs LSTMs over per-frame joint
coordinates in one of two ways. The stacked version has two layers of 512
units on NTU RGB+D. The hierarchical version runs one LSTM per body part
(two arms, two legs and the trunk), concatenates their outputs at each step,
and feeds a second LSTM for the whole body, with 128 units per part and 512
for the body [arxiv:1704.02581 §4.1; §5.2; tab. 2]. The LSTM has peephole
connections from the cell to the gates [arxiv:1704.02581 eq. 1]. In the
paper's ablation on NTU, the stacked stream reaches 66.1% cross-subject and
68.9% cross-view, and the hierarchical stream 67.8% and 70.5% [arxiv:1704.02581
tab. 1]. A second stream reads the joints as an ordered sequence, one time
window per joint. Fusing the two streams gives 68.6% and 71.7%. Adding the 3D
rotation, scaling and shear augmentation gives 71.3% and 79.5%
[arxiv:1704.02581 tab. 1]. The augmentation ranges are not stated
[arxiv:1704.02581 §4.3]. Card: [Modeling Temporal Dynamics and Spatial
Configurations of Actions Using Two-Stream Recurrent Neural Networks](/papers/1704.02581).

### Key-value attention over a knowledge base in an LSTM decoder

The in-car dialogue paper uses an LSTM encoder and an LSTM decoder with
attention over the encoder states [arxiv:1705.05414 §2.1; §2.2]. It adds one
score per knowledge-base key to the decoder's output logits, so the model can
emit a canonical token such as a meeting time. A lookup then swaps that token
for its stored value [arxiv:1705.05414 §2.3; eq. 7–9]. The embedding, hidden
and cell sizes are all 200. Training is token-level cross-entropy with Adam,
and no belief tracker is used [arxiv:1705.05414 §5.2]. On the test split, the
model reports 13.2 BLEU and 48.0 entity F1. Without encoder attention it
reports 10.8 and 40.9, Copy Net reports 11.0 and 37.0, and the rule-based
system 6.6 and 43.8 [arxiv:1705.05414 tab. 3]. Card: [Key-Value Retrieval
Networks for Task-Oriented Dialogue](/papers/1705.05414).

### Bidirectional encoder with attention decoder for speech translation

The speech-translation paper maps French speech to English text with one
attention-based encoder-decoder and no source transcript [arxiv:1612.01744
§1]. The encoder is a hierarchical bidirectional LSTM. It takes 40 MFCCs plus
frame energy per frame, behind two fully connected layers of size 256. Layers
2 and 3 each read every other output of the layer below, so the encoded
sequence is a quarter of the frame count [arxiv:1612.01744 §2; §3]. The
decoder is an LSTM that takes the previous target word. Its attention adds a
convolutional term over the previous step's attention weights, which the
authors say stops the model translating the same part of the signal twice
[arxiv:1612.01744 §2; eq. 6]. Training uses Adam with an initial learning rate
of 0.001, mini-batch size 64 and dropout 0.5 between LSTM layers
[arxiv:1612.01744 §2]. Card: [Listen and Translate: A Proof of Concept for
End-to-End Speech-to-Text Translation](/papers/1612.01744).

On the held-out speaker's test1 set with seven references, the single greedy
model reaches 37.6 BLEU. Beam search, a trigram language model and a 5-model
ensemble raise it to 40.1, 41.8 and 48.8. The Google ASR plus SMT pipeline
reaches 52.7 [arxiv:1612.01744 tab. 3]. On the training speaker the ensemble
reaches 50.4 against 51.9 for the pipeline [tab. 3]. All speech data is
synthetic, and the paper reports no natural-speech result [arxiv:1612.01744
§5]. The card records ensemble weights as not stated, so the ensemble gain
cannot be reproduced exactly from the paper.

## When to use which

| Situation | Use | Why |
|---|---|---|
| Long sequences where decoding must use constant memory per step | mLSTM stack (xLSTM[1:0]) | The paper states linear computation and constant memory in sequence length [arxiv:2405.04517 §2.5]. |
| Training throughput matters and a parallel training form is needed | mLSTM | Its gates do not depend on the previous output, so a parallel form exists [arxiv:2405.04517 app. A.3]. The default released backend builds T x T matrices per head, so training memory grows quadratically; see the card's Paper vs code section. |
| The task needs state tracking that a sequential recurrence can do but a parallel model cannot | sLSTM blocks | On formal-language tasks, Transformers and SSMs without memory mixing cannot solve parity, and xLSTM does best among the tested models [arxiv:2405.04517 §4.1; fig. 4]. |
| Retrieval or sequence tasks where the vanilla LSTM fails | xLSTM, not vanilla LSTM | On Long Range Arena, vanilla LSTM does not beat the random baseline on Retrieval or Pathfinder, while xLSTM scores 0.906 and 0.919 [arxiv:2405.04517 app. B.1.3]. |
| Limited implementation budget and you cannot maintain a matrix memory | Exponential-gated scalar cell only | Exponential gating with a scalar cell reaches 17.70 perplexity versus 13.48 with matrix memory on the 15B ablation [arxiv:2405.04517 tab. 2]. Expect a clear loss. |
| Evaluation beyond the training context length | xLSTM | Trained at context 2048 and tested at 16384, xLSTM[7:1] reaches 8.92 and Llama reaches 337.83 [arxiv:2405.04517 §4.3; fig. 7]. |
| Proprioception-only robot control trained with PPO, where the payload changes in simulation | LSTM controller with a clock input, as in [the locomotion card](/papers/2204.04340) | This is the design the paper reports for dynamic loads. The Carry Pole general policy reaches a 0.8342 pass rate [arxiv:2204.04340 tab. 2]. The card has no feedforward comparison. |
| A new load must be learned and training budget is limited | Bootstrap from a policy trained without the load | Bootstrapped policies reach pass rates of 0.9834 (Carry Pole) and 0.998 (Water Jug), against 0.9578 and 0.976 from scratch [arxiv:2204.04340 tab. 3]. The text says about 100 million fewer samples are needed, and fig. 4 is only a rough reading [arxiv:2204.04340 §IV-C] [fig. 4]. |
| Hardware speed must match simulation | Budget for a gap | Hardware top speed is 3.0 m/s for the base policy against 3.9 m/s in simulation, and 0.8 m/s for Carry Pole against 2.5 m/s [arxiv:2204.04340 §V] [tab. 2]. The paper treats this as a finding, not a fix. |
| Classifying one squeeze of a robot hand from raw sensor series, with no feature design | LSTM on raw series, with SVM on features as the first baseline | On the Barrett Hand, SVM and LSTM both reach 94–99% on the 20-foam set [arxiv:2204.06343 §5.2]. The LSTM is over 75% on the OnRobot RG6 [§5.2]. Transfer across finger configurations fails for the Barrett Hand and the qb SoftHand [§5.4]. |
| Forecasting a teleoperated arm a few seconds ahead from its own joint and force data | CNN feature extractor feeding an LSTM | The quadruplet CNN+LSTM has the lowest forecast error in the paper's table, against 0.065 for the spatial-mean LSTM baseline [arxiv:2402.10220 tab. III]. The paper compares only with that baseline [§IV-E.1]. |
| A physics model exists and its errors depend on terrain | LSTM that corrects the model's forces, with visual features if you have them | The visual model reports about 10% lower mean summed loss than a no-vision model with hindsight features [arxiv:2412.00581 §6.1]. Features help only when they are projected within 0 to 30 m [§6.2]. The paper gives no runtime figures. |
| Skeleton sequences of body joints, with a two-stream RNN as the baseline to beat | Stacked or hierarchical LSTM streams, with the augmentation | On NTU, the fused model reports 71.3% and 79.5%, above the RNN methods in its table, such as part-aware LSTM at 62.9% and 70.3% [arxiv:1704.02581 tab. 3]. No code is released and the augmentation ranges and fusion formula are not stated, so expect to re-derive them. |
| Task-oriented dialogue where answers must contain values from a knowledge base, and no dialogue-state labels exist | Encoder-decoder LSTM with KB-key attention | Entity F1 is 48.0 against 43.8 for the rule-based system and 37.0 for Copy Net [arxiv:1705.05414 tab. 3]. The rule-based system still scores higher on human correctness, 3.96 against 3.70 [tab. 5]. The corpus licence is not stated, so do not redistribute it [§1]. |
| Unlabelled multi-sensor recordings with normal-only training data, and no released code to start from | Convolutional autoencoder with an attention BiLSTM predictor | Author-reported mF1 on PAMAP2 is 0.9639 against 0.8843 for the strongest baseline in the table, ConvLSTM-COMP [arxiv:2107.12626 tab. 2]. On CAP the strongest baseline is OCSVM at 0.9263 [tab. 2]. The learning rate and MMD bandwidth are not stated, so expect to tune them [§4.3]. |
| Sequence predictor for anomaly scoring, where you want to know whether attention is worth keeping | Keep the attention layer | Removing attention lowers author-reported mF1 by 0.038 to 0.069 across the three datasets, from single runs [arxiv:2107.12626 tab. 6]. |
| Speech-to-English text translation with no source transcript, as an early reference on clean, synthetic speech | Bidirectional LSTM encoder with convolutional-attention decoder, ensembled and beam-searched | Test1 BLEU is 48.8 for the 5-model ensemble against 52.7 for the Google ASR plus SMT pipeline, and 37.6 for the greedy single model [arxiv:1612.01744 tab. 3]. Without the ensemble the model falls well behind the pipeline. Speech training takes about 8 hours on one GTX 1070 for 20k steps [arxiv:1612.01744 §4]. |
| Causal, low-latency decoding of continuous movement from surface EMG | LSTM or TCN, with SVR as the frame-wise baseline | Standard session RMSE is 8.0° for the LSTM, 8.4° for the TCN and 11.3° for SVR [arxiv:2412.17991 §3.1]. The two sequential models are within 0.4° of each other, so pick on other grounds. The free-form test on eight subjects gives 9.9° for the LSTM against 10.4° for the TCN [§3.2]. |

All numbers come from the authors. The paper's language-model results use
SlimPajama at 15B tokens, where xLSTM[1:0] reaches 13.43 against 13.70 for
Mamba and 14.25 for Llama [arxiv:2405.04517 tab. 1]. At 300B tokens with a
1.3B model, the figures are 8.89, 9.14 and 9.44 [arxiv:2405.04517 tab. 3].
The paper did not tune the large architectures [arxiv:2405.04517 §5 (vi)].

## Implementation sketch

The shared core is the exponential-gated update with a stabilizer, shown for
the scalar cell. The mLSTM follows the same pattern with a matrix state and
a normalizer division. The sigmoid forget gate is shown; the paper also
allows an exponential one [arxiv:2405.04517 §2.2].

```pseudocode
# Exponential input gate, forget gate in log space, stabilizer  [§2.2]
# State per unit: c (cell), n (normalizer), m (stabilizer)
def exp_gated_step(z, i_pre, f_pre, c_prev, n_prev, m_prev):
    # Forget gate in log space; the input gate is exponential  [§2.2]
    log_f = log_sigmoid(f_pre)
    # Stabilizer keeps every exp() argument at or below zero  [§2.2; app. A.2]
    m = max(log_f + m_prev, i_pre)
    i_gate = exp(i_pre - m)
    f_gate = exp(log_f + m_prev - m)
    # Additive gated memory and normalizer update  [§2.2]
    c = f_gate * c_prev + i_gate * z
    n = f_gate * n_prev + i_gate
    return c, n, m
```

```pseudocode
# Hidden output: output gate times the normalized cell  [§2.2]
# o_gate and z are computed from the input and the previous hidden state
h = o_gate * (c / n)
```

Where the variants differ: the mLSTM replaces `c` with a matrix, uses a
covariance update and divides by a normalizer threshold [arxiv:2405.04517
§2.3]. Released code and paper also differ in several details, listed on
the card under "Paper vs code" [/papers/2405.04517].

The recurrent controller uses a different loop. It takes the robot's sensor
inputs and a clock at each control step and returns joint set points. The
locomotion card has no official code, so the sketch follows its text only
[arxiv:2204.04340 §III-A].

```pseudocode
# One control step of the recurrent policy  [arxiv:2204.04340 §III-A]
# lstm_state is carried from one control step to the next
def policy_step(obs, clock, cmd_vel, lstm_state):
    # Sine and cosine of the gait clock, from its period  [§III-A]
    x = concat(obs, sin(clock), cos(clock), cmd_vel)
    # Two stacked recurrent layers of 128 units  [§III-A]
    h, lstm_state = lstm_2x128(x, lstm_state)
    # Joint set points at 40 Hz; the output head is not described  [§III-A]
    action = output_head(h)
    return action, lstm_state
```

## Contradictions

The xLSTM card and the locomotion card study different uses of the LSTM, a
language model and a locomotion controller. Their numbers are not comparable
and do not conflict. Within the locomotion paper, the text and the table
differ on the Cart force gain. The text gives 60 N more force than the
unloaded robot, while the table gives 40 N [arxiv:2204.04340 §IV-A] [tab. 2].
Within the xLSTM paper, the equations and the released code differ in several
places: the gate inputs, the output gate, the normalizer threshold and the
sLSTM bias initialisation. Follow one source and state which one you chose.
See the card's Paper vs code section [/papers/2405.04517]. The paper also
marks some table values as re-evaluations, which the card notes
[/papers/2405.04517].

The anomaly paper's text and equations disagree on two points. It merges the
BiLSTM directions by sum in the text but writes concatenation in eq. 13, and
it sums attention over all h steps in eq. 17 but over h − 1 steps in its
implementation details [arxiv:2107.12626 §3.4.1; eq. 13; eq. 17; §4.3]. Its
card lists both under Pitfalls, and the page does not choose between them.

The skeleton and dialogue cards add two more tasks, and they do not conflict
with the others either. The skeleton paper's §5.5 swaps the meanings of λ and
τ between its text and the Figure 7 caption, so the window sweep cannot be
read with confidence [arxiv:1704.02581 §5.5; fig. 7].

The robotics cards do not conflict with each other, because they study
different tasks. Their own internal inconsistencies matter for
implementation. In the teleoperation paper, the forecast table labels its
metric RMSE, while the text and figure captions report MSE
[arxiv:2402.10220 tab. III] [arxiv:2402.10220 fig. 13]. Its forecast loss is
also called RMSE in one section and MSE in another [arxiv:2402.10220 §III-G]
[arxiv:2402.10220 §IV-E]. In the off-road paper,
the LSTM initialisation is described as "20 hidden layers", while the
predictor's hidden size is 4 [arxiv:2412.00581 §6]. These may refer to the
same unit, and the paper does not say. In the EMG paper, the unit of the
sequence length is not stated [arxiv:2412.17991 tab. 1].

## Papers

- [xLSTM: Extended Long Short-Term Memory](/papers/2405.04517): introduces exponential gating, sLSTM, mLSTM and the xLSTM block stack, with author-reported language-model results.
- [Modeling Temporal Dynamics and Spatial Configurations of Actions Using Two-Stream Recurrent Neural Networks](/papers/1704.02581): stacked and hierarchical LSTM streams over skeleton joints, fused with a spatial stream that reads the joints as a sequence.
- [Key-Value Retrieval Networks for Task-Oriented Dialogue](/papers/1705.05414): an LSTM encoder-decoder that adds knowledge-base key scores to its output logits, with author-reported dialogue results.
- [Sim-to-Real Learning for Bipedal Locomotion Under Unsensed Dynamic Loads](/papers/2204.04340): uses a two-layer LSTM policy trained with PPO to walk a simulated and a real bipedal robot under dynamic loads.
- [Single-grasp Deformable Object Discrimination: the Effect of Gripper Morphology, Sensing Modalities, and Action Parameters](/papers/2204.06343): compares SVM, k-NN and an LSTM on raw squeeze signals from four robot hands and grippers.
- [Towards Semi-Autonomous Robotic Arm Manipulation Operator Intention Detection from Force Data](/papers/2402.10220): a CNN classifies glovebox tasks and a CNN+LSTM forecasts the next 5 s of arm data.
- [Dynamics Modeling using Visual Terrain Features for High-Speed Autonomous Off-Road Driving](/papers/2412.00581): an LSTM corrects a bicycle-type vehicle model's forces, optionally using DINOv2 terrain features.
- [Online Adaptation for Myographic Control of Natural Dexterous Hand and Finger Movements](/papers/2412.17991): a causal LSTM and TCN decode seven wrist and finger angles from forearm EMG, with online updates.
- [Unsupervised Deep Anomaly Detection for Multi-Sensor Time-Series Signals](/papers/2107.12626): a convolutional autoencoder trained with an attention BiLSTM and an autoregressive predictor, scored by summed loss, with author-reported F1.
- [Listen and Translate: A Proof of Concept for End-to-End Speech-to-Text Translation](/papers/1612.01744): a hierarchical BiLSTM encoder and LSTM decoder with convolutional attention translate French speech directly to English text, with author-reported BLEU on synthetic speech.