Pith. sign in

REVIEW 2 major objections 4 minor 5 cited by

VPWEM: Non-Markovian Visuomotor Policy with Working and Episodic Memory

T0 review · 2 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read VPWEM claims that a fixed-size episodic memory, recursively compressed from evicted observations, gives diffusion policies near-constant-cost access to full episode history and outperforms strong baselines by over 20% on memory-intensive ta

desk verdict A credible memory-compression extension for diffusion policies, but the headline MIKASA margin is built on external baselines and a train/inference subsampling mismatch that the paper never addresses. read the letter →

arxiv 2603.04910 v2 pith:ZTV6LB4H submitted 2026-03-05 cs.RO cs.AIcs.LG

Yuheng Lei , Zhixuan Liang , Hongyuan Zhang , Ping Luo This is my paper
classification cs.ROcs.AIcs.LG
keywords imitationlearningvisuomotorpolicynon-Markoviantasksepisodicmemoryworkingdiffusiontransformercompressorlong-term
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the reason visuomotor policies fail on non-Markovian tasks is not lack of model capacity but lack of a mechanism for long-term memory. It introduces VPWEM, which pairs a short sliding window of recent observations (working memory) with a fixed set of summary tokens produced by a Transformer compressor that recursively ingests observations as they leave the window (episodic memory). Trained end-to-end with a diffusion policy, this compressor gives the policy access to the entire episode history at nearly constant per-step cost. On memory-intensive manipulation tasks the method reports over 20% higher success than strong baselines including vision-language-action models, and an average 5% gain on a mobile manipulation benchmark.

What carries the argument

The central object is the contextual memory compressor: a Transformer encoder block that maintains two caches — a summary cache of its own past output tokens and an observation cache of out-of-window frames. At each step a newly evicted frame is added to the observation cache, and the compressor updates a fixed number M of memory tokens by (1) self-attention over the summary cache and (2) cross-attention over the observation cache, followed by an MLP. These M tokens serve as the episodic memory conditioning for the diffusion policy, while a sliding window of the last L observation tokens provides working memory.

What would settle it

Run the MIKASA ShellGameTouch and RememberColor3 tasks with VPWEM and the baseline VLA models under a single unified codebase, environment version, and seed set; if the average success-rate gap is below 20% or reverses, the central quantitative claim is falsified.

Watch

Extended reading notes

Core claim

VPWEM claims that a small, fixed number of learned episodic-memory tokens, updated recursively as the episode progresses, can carry the task-relevant information from an arbitrarily long observation history into a diffusion policy. The compressor does this by self-attending to a cache of its own past summaries and cross-attending to a cache of historical observation tokens; gradients flow through the compressor into the policy only via the current summary, not through time, so training stays cheap. The result is a policy that conditions on both the recent window and a compressed whole-episode memory, and the paper shows this combination, rather than simply lengthening the context window, is

Load-bearing premise

The headline >20% margin rests on comparing VPWEM's own runs with success rates for the VLA baselines taken directly from other papers, assuming the environments, observation spaces, seeds, and evaluation protocol are identical across those sources.

Editorial extensions

If this is right

  • VPWEM reports success-rate gains of more than 20% over state-of-the-art baselines on memory-intensive MIKASA manipulation tasks.
  • It reports an average 5% improvement over baselines on the MoMaRT mobile manipulation benchmark.
  • Per-step training and inference cost stays nearly constant as episode length grows, because only the evicted frame is compressed at each step.
  • On an almost-Markovian benchmark, adding the memory module neither hurts nor helps, suggesting the mechanism is compatible with standard policies.
  • The memory module is small (~2.2M parameters) and can be attached to different diffusion backbones, improving both.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the compressor's attention weights are inspected, they may reveal which past frames a task actually depends on, offering a diagnostic for non-Markovian structure in demonstrations.
  • The same episodic-memory design could be applied to other policy families (e.g., flow-based or autoregressive action generation) without retraining the backbone, since it only changes the conditioning context.
  • A potential failure mode not explored in the paper: when the task requires sparse, precise recall (like an exact color or position seen once early on), a fixed-size summary could bottleneck the information; testing this with a task where the relevant memory is far back and easy to verbalize would be a sharp check.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. This paper proposes VPWEM, a non-Markovian visuomotor policy that augments diffusion-based policies with two memory modules: a sliding-window working memory and a learned episodic memory. The episodic memory is produced by a Transformer compressor that recursively converts out-of-window observation tokens into a fixed number of summary tokens via self-attention over a summary cache and cross-attention over an observation cache. The compressor is trained end-to-end with the policy via a behavior-cloning loss. The authors instantiate the idea on Diffusion Policy and MaIL, evaluate on MIKASA, MoMaRT, and Robomimic, and report large gains on memory-intensive tasks, near-constant compute, and on-par performance on Markovian tasks.

Significance. If the central mechanism performs as claimed, VPWEM offers a practical way to give diffusion policies bounded-memory access to long episode histories, which is useful for non-Markovian manipulation and mobile manipulation. The architecture is clearly specified, the code is available, and the ablations over context length, compressor layers, cache size, and cache management are consistent with the design narrative. The internal comparisons against DP and MaIL are run under the same harness and show consistent gains, and the efficiency analysis in Table II is informative. However, two load-bearing issues—the train/inference subsampling mismatch in the compressor and the reliance on externally sourced VLA baseline numbers—mean the headline claims are not yet fully supported.

major comments (2)
  1. [Section IV-C, Eq. (3), Table I] During training, the compressor's input is a random subsample of out-of-window observations (subsample ratio 5 on MIKASA, 20 elsewhere), so each observation-cache entry is ~5-20 timesteps apart. At inference, every out-of-window token is pushed into the FIFO cache, so cross-attention in Eq. (3) sees dense, adjacent frames. The compressor is therefore optimized on a different input distribution than the one it encounters at deployment. This makes the central claim that the policy has 'near-constant-cost access to full episode history through episodic memory' unsupported: the episodic summary may be acting as a regularizer. No ablation varies the subsample ratio or evaluates with matched subsampling at inference. Please provide an ablation (e.g., train with ratio 1, or evaluate with the same subsampling as training) to show the mechanism transfers.
  2. [Section V-B, Figure 5] The abstract's headline result (>20% improvement over VLA baselines) uses success rates for Octo, OpenVLA, pi0, SpatialVLA, CronusVLA, and MemoryVLA 'taken directly from [32], [50]'. These numbers were not produced under the same evaluation protocol, environment version, or seeds as the VPWEM runs. Unless the authors verify that the conditions are identical, the margin may reflect protocol differences rather than algorithmic gain. At minimum, the paper should state the exact source numbers and any differences in evaluation; ideally, rerun at least one VLA baseline under the same harness.
minor comments (4)
  1. [Section IV-B, Eq. (2)-(3)] The shape of the summary cache \bar{C}_q is given as R^{S×M×D}, but self-attention expects a 2D key/value matrix. Please clarify how the cache is flattened or reshaped before the attention computation.
  2. [Section IV-C, Training] The procedure for constructing the subsampled input to the compressor is underspecified: how are segments defined relative to the episode, and how is the maximum cache size enforced during training? Please give a precise algorithm.
  3. [Figure 7] The ablation bars are difficult to read without numerical values; please include the mean±std success rates in the text or as a table.
  4. [Section IV, first paragraph] Typo: 'working meory' should be 'working memory'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the memory compressor is trained end-to-end (Eq. 6) and the headline numbers are external-benchmark comparisons, not constructions from the method's own inputs.

full rationale

VPWEM's derivation chain is empirical, not analytic. The compressor (Eqs. 1-5) is a learned module optimized with the behavior-cloning loss Eq. (6); its output e_tau is not a fitted constant later renamed a prediction. The MIKASA comparison uses baseline numbers 'taken directly from [32], [50]' (Sec. V-B), which is external evidence, and no load-bearing claim rests on a self-citation (the only plausible self-citation, [56], appears in related work and is not used to justify VPWEM). The skeptic's train/inference subsampling mismatch (Sec. IV-C: training subsamples out-of-window tokens; inference compresses all of them) is a real distribution-shift risk, but it is not circular: it does not make any reported success rate equal to a training input by construction. Hyperparameters L2S2, cache size 8, and subsample ratio were chosen by ablation (Sec. V-C), which is tuning, not a fitted-input prediction. The central claim therefore has independent empirical content and is not forced by definition or by self-citation.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a small set of architecture choices (fixed-size memory tokens, cache size, subsample ratio) and on benchmark comparability; none of these is derived from first principles, so they are tabulated as free parameters/axioms rather than hidden in the equations.

free parameters (6)
  • Working memory window length L = 2
    Ablated in Figure 7; chosen as L2S2 for main runs (Table I). Not derived from theory.
  • Episodic memory token count M = 2
    Ablated; more tokens sometimes help then degrade; L2S2 chosen for main runs.
  • Compressor layers N = 2
    Ablation shows a trend of improvement with a larger compressor; N=2 is used, likely as a cost-performance tradeoff.
  • Maximum cache size S = 8
    Ablation shows gains saturate above 8; chosen as default.
  • History subsample ratio = 5 (MIKASA), 20 (otherwise)
    Training randomly samples one frame per segment; set by hand, not derived from theory.
  • Memory token dropout = 0.3
    Regularization hyperparameter in Table I, chosen by hand.
assumptions (5)
  • domain assumption A POMDP policy that conditions on full history h_t = o_{<=t} is necessary for non-Markovian tasks.
    Sec III frames the problem as POMDP and the central motivation is that short-context policies fail on tasks requiring long-term dependencies.
  • ad hoc to paper Fixed-size memory tokens can capture all decision-relevant information from arbitrarily long out-of-window histories.
    Sec IV-B; capacity assumption not supported by an information-theoretic argument, only by ablations.
  • ad hoc to paper Detaching f_tau and q_n,tau from the graph before caching still permits useful end-to-end learning of compression.
    Sec IV-C training paragraph; gradient stopping is a design choice whose sufficiency is empirically assumed.
  • ad hoc to paper Random subsampling of history during training is a faithful substitute for sequential compression during inference.
    Sec IV-C and Figure 2; train/inference distribution mismatch is not analyzed.
  • domain assumption Baseline scores quoted from MIKASA reports [32], [50] are directly comparable to the authors' runs.
    Sec V-B; the >20% claim depends on comparability of environment versions, seeds, and evaluation protocols.
invented entities (1)
  • Episodic memory tokens e_tau (and summary/observation caches)
    purpose: Represent the compressed long-term history for conditioning the diffusion denoiser.
    Learned latent representations introduced by the paper; no external measurable property beyond task success, so no independent falsifiable handle is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VPWEM: Non-Markovian Visuomotor Policy with Working and Episodic Memory." pith.science (2026). https://pith.science/paper/ZTV6LB4H

@misc{pith2026260304910,
  author       = {Pith},
  title        = {Pith review of: VPWEM: Non-Markovian Visuomotor Policy with Working and Episodic Memory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZTV6LB4H}},
  note         = {Machine review of arXiv:2603.04910}
}
read the original abstract

Imitation learning from human demonstrations has achieved significant success in robotic control, yet most visuomotor policies still condition on single-step observations or short-context histories, making them struggle with non-Markovian tasks that require long-term memory. Simply enlarging the context window incurs substantial computational and memory costs and encourages overfitting to spurious correlations, leading to catastrophic failures under distribution shift and violating real-time constraints in robotic systems. By contrast, humans can compress important past experiences into long-term memories and exploit them to solve tasks throughout their lifetime. In this paper, we propose VPWEM, a non-Markovian visuomotor policy equipped with working and episodic memories. VPWEM retains a sliding window of recent observation embeddings as short-term working memory, and introduces a Transformer-based contextual memory compressor that recursively converts out-of-window observations into a fixed number of episodic memory embeddings. The compressor uses self-attention over a cache of past summary embeddings and cross-attention over a cache of historical observations, and is trained jointly with the policy. We instantiate VPWEM on diffusion policies to exploit both short-term and episode-wide information for action generation with nearly constant memory and computation per step. Experiments demonstrate that VPWEM outperforms state-of-the-art baselines including diffusion policies and vision-language-action (VLA) models by more than 20\% on the memory-intensive manipulation tasks in MIKASA and achieves an average 5\% improvement on the mobile manipulation benchmark MoMaRT. Code is available at https://github.com/HarryLui98/code_vpwem.

Figures

Figures reproduced from arXiv: 2603.04910 by the authors.

Figure 1
Figure 1. Comparison between VPWEM and existing methods: (a) Policies [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the VPWEM framework. (a) Policy architecture. Modules in green are learnable components, including the multi-modal encoder, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Three benchmarks in our experiments. Benchmark. We evaluate our method across three benchmarks, as shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance on MoMaRT benchmark. TABLE II COMPARISON OF COMPUTATIONAL AND STORAGE COSTS ON THE UNLOAD-SUBOPTIMAL TASK IN THE MOMART BENCHMARK Methods DP-PTP DP-VPWEM Context Length 4 8 16 32 64 128 L2S2 Model Size (M) 50.67 50.67 50.68 50.69 50.70 50.73 52.98 GPU Memor…
Figure 6
Figure 6. Figure 6: Performance on Robomimic benchmark. MoMaRT benchmark. While the DP-PTP model size remains nearly constant, both training and inference time increase substantially as the context length grows, whereas the success rate improves only marginally and even degrades at the co…
Figure 5
Figure 5. Figure 5: Performance on MIKASA benchmark. merely scaling the policy while conditioning on a single frame is insufficient to solve such non-Markovian tasks [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 7
Figure 7. Figure 7: Ablation studies on Long-Horizon Square task. L1S1 (i.e., one long-term episodic token and one short-term working memory token) variant of our proposed method already achieves 65% success rate, surpassing all previous baselines. Performance generally improves as the nu…

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. FM-VLA: Force-based Memory for Vision-Language-Action Models in Contact-Rich Manipulation

    cs.RO 2026-07 conditional novelty 7.0 of 10

    A vision-language-action robot policy that stores compressed wrist-force histories as memory tokens can count contact events that are visually ambiguous, achieving 83.3% average success across three tasks.

  2. ChainVLA: Chaining Vision-Language-Action Queries through a Unified Execution State for Long-Horizon Manipulation

    cs.RO 2026-08 conditional novelty 6.0 of 10

    ChainVLA chains vision-language-action queries by carrying both retrieved task progress and the previous prediction's unexecuted suffix into each new action horizon, reaching 62.8% on RMBench and 98.8% on LIBERO.

  3. EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    EventVLA introduces foundational visual anchors and a Keyframe Evidence Memory module that predicts future keyframe probabilities from VLA embeddings to improve long-horizon task success by an average of 40% on 17 sim...

  4. $\mu$VLA: On Recurrent Memory for Partially Observable Manipulation in VLA Models

    cs.LG 2026-06 unverdicted novelty 6.0 of 10

    Adding recurrent memory tokens to VLA models raises success rates on partially observable manipulation tasks from 0.42 to 0.84 on training and 0.07 to 0.23 on held-out tasks while preserving performance under full obs...

  5. EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies

    cs.CV 2026-06 unverdicted novelty 5.0 of 10

    EventVLA uses a foresight-driven Keyframe Evidence Memory module to selectively store sparse task-critical visual events predicted from VLA latent embeddings, reporting +40% average success rate gains on 17 simulation...

Reference graph

Works this paper leans on

58 extracted references · 10 linked inside Pith · cited by 4 Pith papers

  1. [32]

    Memory, benchmark & robots: A benchmark for solving complex tasks with reinforcement learning,

    E. Cherepanov, N. Kachaev, A. K. Kovalev, and A. I. Panov, “Memory, benchmark & robots: A benchmark for solving complex tasks with reinforcement learning,”arXiv preprint arXiv:2502.10550, 2025

    arXiv 2025

  2. [50]

    Memoryvla: Perceptual-cognitive memory in vision-language-action models for robotic manipulation,

    H. Shi, B. Xie, Y . Liu, L. Sun, F. Liu,et al., “Memoryvla: Perceptual-cognitive memory in vision-language-action models for robotic manipulation,”arXiv preprint arXiv:2508.19236, 2025

    arXiv 2025

  3. [1]

    Recent advances in robot learning from demonstration,

    H. Ravichandar, A. S. Polydoros, S. Chernova, and A. Billard, “Recent advances in robot learning from demonstration,”Annual review of control, robotics, and autonomous systems, vol. 3, no. 1, pp. 297– 330, 2020

    2020

  4. [2]

    A review of robot learning for manipulation: Challenges, representations, and algorithms,

    O. Kroemer, S. Niekum, and G. Konidaris, “A review of robot learning for manipulation: Challenges, representations, and algorithms,” Journal of machine learning research, vol. 22, no. 30, pp. 1–82, 2021

    2021

  5. [3]

    R3M: A universal visual representation for robot manipulation,

    S. Nair, A. Rajeswaran, V . Kumar, C. Finn, and A. Gupta, “R3M: A universal visual representation for robot manipulation,” inConference on Robot Learning, pp. 892–909, PMLR, 2023

    2023

  6. [4]

    Real-world robot learning with masked visual pre-training,

    I. Radosavovic, T. Xiao, S. James, P. Abbeel, J. Malik, and T. Darrell, “Real-world robot learning with masked visual pre-training,” in Conference on Robot Learning, pp. 416–426, PMLR, 2023

    2023

  7. [5]

    RT-2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid,et al., “RT-2: Vision-language-action models transfer web knowledge to robotic control,” inConference on Robot Learning, pp. 2165–2183, PMLR, 2023

    2023

  8. [6]

    Multimodal diffusion transformer: Learning versatile behavior from multimodal goals,

    M. Reuss, Ö. E. Ya ˘gmurlu, F. Wenzel, and R. Lioutikov, “Multimodal diffusion transformer: Learning versatile behavior from multimodal goals,”arXiv preprint arXiv:2407.05996, 2024

    arXiv 2024

Show all 58 references
  1. [7]

    Openvla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong,et al., “Openvla: An open-source vision-language-action model,” in8th Annual Conference on Robot Learning, 2024

    2024

  2. [8]

    π 0: A vision-language- action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter,et al., “π 0: A vision-language- action flow model for general robot control,”Robotics: Science and Systems, 2025

    2025

  3. [9]

    Octo: An open-source generalist robot policy,

    D. Ghosh, H. R. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, J. Luo,et al., “Octo: An open-source generalist robot policy,” inRobotics: Science and Systems, 2024

    2024

  4. [10]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,”The International Journal of Robotics Research, 2024

    2024

  5. [11]

    3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,”Robotics: Science and Systems, 2024

    2024

  6. [12]

    RDT-1B: a diffusion foundation model for bimanual manipulation,

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu, “RDT-1B: a diffusion foundation model for bimanual manipulation,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  7. [13]

    RT- 1: Robotics transformer for real-world control at scale,

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsu,et al., “RT- 1: Robotics transformer for real-world control at scale,”Robotics: Science and Systems, 2023

    2023

  8. [14]

    Otter: A vision-language-action model with text-aware visual feature extraction,

    H. Huang, F. Liu, L. Fu, T. Wu, M. Mukadam, J. Malik, K. Goldberg, and P. Abbeel, “Otter: A vision-language-action model with text-aware visual feature extraction,” inForty-second International Conference on Machine Learning, 2025

    2025

  9. [15]

    Bidirectional decoding: Improving action chunking via guided test- time sampling,

    Y . Liu, J. I. Hamid, A. Xie, Y . Lee, M. Du, and C. Finn, “Bidirectional decoding: Improving action chunking via guided test- time sampling,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  10. [16]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    2017

  11. [17]

    X-il: Exploring the design space of imitation learning policies,

    X. Jia, A. Donat, X. Huang, X. Zhao, D. Blessing, H. Zhou, H. A. Wang, H. Zhang, Q. Wang, R. Lioutikov,et al., “X-il: Exploring the design space of imitation learning policies,”arXiv preprint arXiv:2502.12330, 2025

    2025 arXiv

  12. [18]

    Transformers are rnns: Fast autoregressive transformers with linear attention,

    A. Katharopoulos, A. Vyas, N. Pappas, and F. Fleuret, “Transformers are rnns: Fast autoregressive transformers with linear attention,” in International conference on machine learning, pp. 5156–5165, PMLR, 2020

    2020

  13. [19]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inFirst Conference on Language Modeling, 2023

    2023

  14. [20]

    Transformers are ssms: Generalized models and efficient algorithms through structured state space duality,

    T. Dao and A. Gu, “Transformers are ssms: Generalized models and efficient algorithms through structured state space duality,” in International Conference on Machine Learning, pp. 10041–10071, PMLR, 2024

    2024

  15. [21]

    Mail: Improving imitation learning with selective state space models,

    X. Jia, Q. Wang, A. Donat, B. Xing, G. Li, H. Zhou, O. Celik, D. Blessing, R. Lioutikov, and G. Neumann, “Mail: Improving imitation learning with selective state space models,” in8th Annual Conference on Robot Learning, 2024

    2024

  16. [22]

    Causal confusion in imitation learning,

    P. De Haan, D. Jayaraman, and S. Levine, “Causal confusion in imitation learning,”Advances in neural information processing systems, vol. 32, 2019

    2019

  17. [23]

    Fighting copycat agents in behavioral cloning from observation histories,

    C. Wen, J. Lin, T. Darrell, D. Jayaraman, and Y . Gao, “Fighting copycat agents in behavioral cloning from observation histories,” Advances in Neural Information Processing Systems, vol. 33, pp. 2564–2575, 2020

    2020

  18. [24]

    Fighting fire with fire: Avoiding dnn shortcuts through priming,

    C. Wen, J. Qian, J. Lin, J. Teng, D. Jayaraman, and Y . Gao, “Fighting fire with fire: Avoiding dnn shortcuts through priming,” inInternational Conference on Machine Learning, pp. 23723–23750, PMLR, 2022

    2022

  19. [25]

    Regularized behavior cloning for blocking the leakage of past action information,

    S. Seo, H. Hwang, H. Yang, and K.-E. Kim, “Regularized behavior cloning for blocking the leakage of past action information,”Advances in Neural Information Processing Systems, vol. 36, pp. 2128–2153, 2023

    2023

  20. [26]

    Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory.,

    J. L. McClelland, B. L. McNaughton, and R. C. O’Reilly, “Why there are complementary learning systems in the hippocampus and neocortex: insights from the successes and failures of connectionist models of learning and memory.,”Psychological review, vol. 102, no. 3, p. 419, 1995

    1995

  21. [27]

    Adapting language models to compress contexts,

    A. Chevalier, A. Wettig, A. Ajith, and D. Chen, “Adapting language models to compress contexts,” in2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, pp. 3829– 3846, Association for Computational Linguistics (ACL), 2023

    2023

  22. [28]

    Artificial hippocampus networks for efficient long-context modeling,

    Y . Fang, W. Yu,et al., “Artificial hippocampus networks for efficient long-context modeling,”arXiv preprint arXiv:2510.07318, 2025

    2025

  23. [29]

    Memagent: Reshaping long- context llm with multi-conv rl-based memory agent,

    H. Yu, T. Chen, J. Feng,et al., “Memagent: Reshaping long- context llm with multi-conv rl-based memory agent,”arXiv preprint arXiv:2507.02259, 2025

    2025 arXiv

  24. [30]

    Ma-lmm: Memory-augmented large multimodal model for long-term video understanding,

    B. He, H. Li, Y . K. Jang, M. Jia, X. Cao, A. Shah, A. Shrivastava, and S.-N. Lim, “Ma-lmm: Memory-augmented large multimodal model for long-term video understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13504– 13514, 2024

    2024

  25. [31]

    Memory consolidation enables long-context video understanding,

    I. Balazevic, Y . Shi, P. Papalampidi, R. Chaabouni, S. Koppula, and O. J. Henaff, “Memory consolidation enables long-context video understanding,” inInternational Conference on Machine Learning, pp. 2527–2542, PMLR, 2024

    2024

  26. [33]

    Error-aware imitation learning from teleoperation data for mobile manipulation,

    J. Wong, A. Tung, A. Kurenkov, A. Mandlekar, L. Fei-Fei, S. Savarese, and R. Martín-Martín, “Error-aware imitation learning from teleoperation data for mobile manipulation,” inConference on Robot Learning, pp. 1367–1378, PMLR, 2022

    2022

  27. [34]

    What matters in learning from offline human demonstrations for robot manipulation,

    A. Mandlekar, D. Xu, J. Wong, S. Nasiriany, C. Wang, R. Kulkarni, L. Fei-Fei, S. Savarese, Y . Zhu, and R. Martín-Martín, “What matters in learning from offline human demonstrations for robot manipulation,” inConference on Robot Learning, pp. 1678–1690, PMLR, 2022

    2022

  28. [35]

    Advancing transformer architecture in long-context large language models: A comprehensive survey,

    Y . Huang, J. Xu, J. Lai, Z. Jiang, T. Chen, Z. Li, Y . Yao, X. Ma, L. Yang, H. Chen,et al., “Advancing transformer architecture in long-context large language models: A comprehensive survey,”arXiv preprint arXiv:2311.12351, 2023

    2023 arXiv

  29. [36]

    A comprehensive survey on long context language modeling,

    J. Liu, D. Zhu, Z. Bai, Y . He, H. Liao, H. Que, Z. Wang, C. Zhang, G. Zhang, J. Zhang,et al., “A comprehensive survey on long context language modeling,”arXiv preprint arXiv:2503.17407, 2025

    2025

  30. [37]

    Roformer: Enhanced transformer with rotary position embedding,

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “Roformer: Enhanced transformer with rotary position embedding,” Neurocomputing, vol. 568, p. 127063, 2024

    2024

  31. [38]

    Longformer: The long- document transformer,

    I. Beltagy, M. E. Peters, and A. Cohan, “Longformer: The long- document transformer,”arXiv preprint arXiv:2004.05150, 2020

    2004 arXiv

  32. [39]

    Transformer-XL: Attentive language models beyond a fixed-length context,

    Z. Dai, Z. Yang, Y . Yang, J. Carbonell, Q. V . Le, and R. Salakhutdinov, “Transformer-XL: Attentive language models beyond a fixed-length context,”arXiv preprint arXiv:1901.02860, 2019

    1901 arXiv

  33. [40]

    Memorizing transformers,

    Y . Wu, M. N. Rabe, D. Hutchins, and C. Szegedy, “Memorizing transformers,” inInternational Conference on Learning Representations, 2022

    2022

  34. [41]

    Recurrent memory transformer,

    A. Bulatov, Y . Kuratov, and M. Burtsev, “Recurrent memory transformer,”Advances in Neural Information Processing Systems, vol. 35, pp. 11079–11091, 2022

    2022

  35. [42]

    xlstm: Extended long short-term memory,

    M. Beck, K. Pöppel, M. Spanring, A. Auer, O. Prudnikova, M. Kopp, G. Klambauer, J. Brandstetter, and S. Hochreiter, “xlstm: Extended long short-term memory,”Advances in Neural Information Processing Systems, vol. 37, pp. 107547–107603, 2024

    2024

  36. [43]

    Packing input frame context in next-frame prediction models for video generation,

    L. Zhang and M. Agrawala, “Packing input frame context in next-frame prediction models for video generation,”arXiv preprint arXiv:2504.12626, 2025

    2025

  37. [44]

    Mtil: Encoding full history with mamba for temporal imitation learning,

    Y . Zhou, Y . Lin, F. Peng,et al., “Mtil: Encoding full history with mamba for temporal imitation learning,”IEEE Robotics and Automation Letters, 2025

    2025

  38. [45]

    In-context imitation learning via next-token prediction,

    L. Fu, H. Huang, G. Datta, L. Y . Chen, W. C.-H. Panitch, F. Liu, H. Li, and K. Goldberg, “In-context imitation learning via next-token prediction,”arXiv preprint arXiv:2408.15980, 2024

    2024 arXiv

  39. [46]

    Action tokenizer matters in in-context imitation learning,

    A. D. Vuong, M. N. Vu, D. An, and I. Reid, “Action tokenizer matters in in-context imitation learning,”arXiv preprint arXiv:2503.01206, 2025

    2025

  40. [47]

    Learning long-context diffusion policies via past-token prediction,

    M. Torne, A. Tang, Y . Liu, and C. Finn, “Learning long-context diffusion policies via past-token prediction,” in9th Annual Conference on Robot Learning, 2025

    2025

  41. [48]

    Cronusvla: Transferring latent motion across time for multi-frame prediction in manipulation,

    H. Li, S. Yang, Y . Chen, Y . Tian,et al., “Cronusvla: Transferring latent motion across time for multi-frame prediction in manipulation,”arXiv preprint arXiv:2506.19816, 2025

    2025

  42. [49]

    Contextvla: Vision-language-action model with amortized multi-frame context,

    H. Jang, S. Yu, H. Kwon, H. Jeon, Y . Seo, and J. Shin, “Contextvla: Vision-language-action model with amortized multi-frame context,” arXiv preprint arXiv:2510.04246, 2025

    2025

  43. [51]

    Implicit behavioral cloning,

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mordatch, and J. Tompson, “Implicit behavioral cloning,” inConference on robot learning, pp. 158–168, PMLR, 2022

    2022

  44. [52]

    Cleandiffuser: An easy-to-use modularized library for diffusion models in decision making,

    Z. Dong, Y . Yuan, J. Hao, F. Ni, Y . Ma, P. Li, and Y . Zheng, “Cleandiffuser: An easy-to-use modularized library for diffusion models in decision making,”Advances in Neural Information Processing Systems, vol. 37, pp. 86899–86926, 2024

    2024

  45. [53]

    Language control diffusion: Efficiently scaling through space, time, and tasks,

    E. Zhang, Y . Lu, S. Huang, W. Y . Wang, and A. Zhang, “Language control diffusion: Efficiently scaling through space, time, and tasks,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  46. [54]

    Imitating human behaviour with diffusion models,

    T. Pearce, T. Rashid, A. Kanervisto, D. Bignell, M. Sun, R. Georgescu, S. V . Macua, S. Z. Tan, I. Momennejad, K. Hofmann,et al., “Imitating human behaviour with diffusion models,” inInternational Conference on Learning Representations, 2023

    2023

  47. [55]

    Flowpolicy: Enabling fast and robust 3d flow-based policy via consistency flow matching for robot manipulation,

    Q. Zhang, Z. Liu, H. Fan, G. Liu, B. Zeng, and S. Liu, “Flowpolicy: Enabling fast and robust 3d flow-based policy via consistency flow matching for robot manipulation,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, pp. 14754–14762, 2025

    2025

  48. [56]

    Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language- action policies,

    Z. Liang, Y . Li, T. Yang, C. Wu,et al., “Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language- action policies,”arXiv preprint arXiv:2508.20072, 2025

    2025 arXiv

  49. [57]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  50. [58]

    Spatialvla: Exploring spatial representations for visual-language-action model,

    D. Qu, H. Song, Q. Chen, Y . Yao, X. Ye,et al., “Spatialvla: Exploring spatial representations for visual-language-action model,” arXiv preprint arXiv:2501.15830, 2025

    2025 arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.