Pith. sign in

REVIEW 3 major objections 5 minor 43 references

Naively scaling Diffusion Policy context length is not as brittle as prior work claimed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 13:48 UTC pith:5KMRH4Q7

load-bearing objection Solid multi-factor empirical corrective: naive long-context Diffusion Policy works with UNet+Cross-Attention and enough data; variable-history training is a practical low-data fix. Single-seed best-checkpoint selection is the real soft spot, not a load-bearing collapse. the 3 major comments →

arxiv 2606.16447 v2 pith:5KMRH4Q7 submitted 2026-06-15 cs.RO cs.AI

Training and Evaluating Diffusion Policies with Long Context Lengths

Abhinav Agarwal , Adam Wei , Taylan Kargin , Michael Zeng , Cole Becker , Arif Kerem Dayi , Pablo Parrilo , Asuman Ozdaglar
show 1 more author
Russ Tedrake
This is my paper
classification cs.RO cs.AI
keywords diffusion policyimitation learninglong contextrobotic manipulationcross-attentionvariable history trainingsample complexity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Robotic imitation policies usually see only a short slice of recent camera frames, so they cannot remember earlier events and often loop on the same failed motion. This paper asks whether simply feeding much longer observation histories into Diffusion Policies really fails as badly as recent papers have said. Across five tasks that differ in how stable the manipulation is and how much memory they need, and across three data budgets per task, the authors find that with a UNet denoiser conditioned by cross-attention, long histories often work well once you have a normal amount of demonstration data. They also introduce a curriculum that trains one policy on many history lengths at once, which closes much of the remaining gap when data is scarce, and they re-examine an earlier past-action prediction trick, finding that freezing the vision encoder is doing more work than previously advertised.

Core claim

With an appropriate conditioning method and denoising backbone (UNet plus cross-attention), single-task Diffusion Policies achieve high success rates on many robotic manipulation tasks in the usual data regime even when context length is naively scaled to long horizons (up to tens of frames, and in one hardware case 92). The sample complexity of long-context learning is driven largely by how locally stable the manipulation primitive is, not by history length alone.

What carries the argument

UNet+Cross-Attention conditioning, which keeps per-timestep observation tokens separate and cross-attends them into the UNet rather than collapsing them through FiLM; plus Variable History Training, which samples training windows from a curriculum of shorter and longer context lengths so the same policy learns both short-horizon control and useful memory.

Load-bearing premise

That ranking architectures and curricula from single training runs (best closed-loop checkpoint, Wilson intervals on 200 rollouts) is enough, even though multi-seed training variance is not measured.

What would settle it

Retrain the same UNet+Cross-Attention, UNet+FiLM, and DiT configurations on push-and-return and square at N/2 and N with several random seeds; if long-context Cross-Attention no longer outperforms FiLM or if naive long context still collapses relative to short context across seeds, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Default single-task Diffusion Policy baselines should prefer UNet with cross-attention when context is long and data is limited.
  • Long-context failure is often a manipulation-skill learning problem, not a pure memory problem: policies that finish the motion usually also track history.
  • Variable History Training can be used without first measuring whether the dataset is large enough for naive scaling.
  • Prior criticism of naive history scaling was partly driven by architecture choice (especially DiT) and missing data-scale ablations.
  • Past-action prediction alone is not a reliable fix; freezing a short-context vision encoder contributes materially to reported gains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If long-context failure is mainly covariate shift from overfitting, similar multi-length curricula may help other visuomotor policy classes beyond diffusion.
  • Locally stable prehensile tasks may systematically understate the data cost of long memory; hard contact and non-prehensile tasks remain the stricter testbed.
  • Inference-time adaptive context length (short when local control is enough, long when memory is needed) is a natural next step for policies already trained on variable histories.
  • Hardware success with 100 demos and To=92 suggests some memory-heavy kitchen tasks tolerate rough grasps enough that naive scaling can already be practical.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper systematically studies how Diffusion Policy performance changes as observation context length To is scaled from short to long (up to 80–92). Across five tasks that vary in local manipulation stability and memory demand, three data regimes (N/2, N, 2N), and three architectures (UNet+FiLM, UNet+Cross-Attention, DiT), the authors report that naive long-context scaling is not catastrophically brittle when UNet+Cross-Attention is used and data is at usual single-task scale. They introduce Variable History Training (Algorithm 1), a curriculum over multiple context lengths that improves low-data long-context success while preserving high-data performance, and re-examine past-token prediction with frozen encoders. Supporting metrics (task success, manipulation completion, contextual success) and limited hardware results on marshmallows (To=92) are provided.

Significance. If the empirical picture holds, the work revises a widely cited premise in long-context imitation learning: that naive history scaling fails and therefore requires heuristic compression, VLM filtering, or auxiliary losses. The breadth of the sweep (~200 policies), the introduction of memory-demanding tasks with interpretable success decompositions, the architecture ablation that isolates UNet+Cross-Attention as a useful inductive bias, and the simple curriculum (Algorithm 1) are concrete contributions that practitioners can adopt. The re-evaluation of past-token prediction and encoder freezing is also useful. Strengths include clear success metrics, public website/code intent, and explicit discussion of failure modes (overfitting vs. memory tracking).

major comments (3)
  1. Appendix B.3 and D.3: each reported success rate is from a single training run, with the best of many checkpoints chosen by closed-loop success (including early/late snapshots). The paper itself notes that longer-context policies in the N/2 regime often converge earlier and that training/validation loss can fall while closed-loop success falls (Fig. 12). Wilson intervals on 200 rollouts capture only evaluation noise. Without multi-seed variance, the architecture ranking in Fig. 7 and the low-data gains of progressive+short variable history in Fig. 8 could be inflated by lucky early checkpoints. At least 2–3 seeds on the key comparisons (UNet+xAttn vs FiLM at long To; variable history vs naive at N/2) are needed for the central claim against prior “naive scaling is brittle” literature to be secure.
  2. Section 4.1 / Fig. 7 and Appendix E.1: DiT is shown to fail, but the authors substantially alter the original DiT design (more layers, non-causal attention within the action chunk, extra sampling steps) while still reporting catastrophic failure on Drake tasks. The claim that DiT “should clearly not be chosen as a baseline” is therefore only partially supported; either a closer reproduction of Torne et al.’s hyperparameters or an explicit statement that even a capacity-matched, non-causal DiT fails is required so that the architecture comparison fairly represents prior long-context work.
  3. Section 3.2 and Appendix D.2: the claim that sample complexity of long-context learning is primarily dictated by local stability of the manipulation primitive rests on a small set of tasks (lift/grasp-and-return vs push-T/square/push-and-return) and a single contact-offset diagnostic (Table 4). The axiom is plausible but not yet isolated from other confounds (action dimensionality, data-generation method, observability of phase). A controlled ablation that holds memory structure fixed while varying only local stability would strengthen this load-bearing interpretation.
minor comments (5)
  1. Algorithm 1: the notation Tpast_p(m)=min{m,Tpast_p} and the role of ρ_i are clear, but the recommended default (progressive+short vs random sprinkle+full) is only stated in prose in Appendix E.2; a short decision rule in the main text would help practitioners.
  2. Figure 1 caption and abstract claim “first study to investigate context length … at this level of detail”; Mark et al. [9] already provide limited data-scaling ablations. Soften the priority claim or cite their agreement more precisely.
  3. Table 1: parameter counts for FiLM grow with To while xAttn stays fixed; the main comparison keeps UNet channel widths fixed rather than total parameters. Appendix E.1 partially addresses this, but a one-sentence pointer in Section 4.1 would avoid confusion.
  4. Typos / wording: “auxilliary” (Section 1.1), “datagrams” (Fig. 1 / Algorithm 1) is nonstandard for trajectory segments; “insentive” (Appendix B.3).
  5. Hardware: marshmallows uses only 20 trials and the latest checkpoint; state this limitation more prominently when claiming high success at To=92.

Circularity Check

0 steps flagged

Empirical benchmarking paper with no circular derivation: success rates are measured on held-out rollouts, not forced by fitted constants or self-definitional loops.

full rationale

The paper's central claims (naive long-context scaling is not catastrophic with UNet+Cross-Attention; variable-history training reduces sample complexity in the low-data regime; past-token prediction success is partly driven by encoder freezing) are established by training ~200 policies and reporting closed-loop success rates on held-out rollouts (200 trials, Wilson intervals). There is no mathematical derivation in which a quantity is defined from data and then re-presented as a prediction of that same quantity. Architecture comparisons (UNet+FiLM vs UNet+Cross-Attention vs DiT), data-scale ablations, and the variable-history curriculum (Algorithm 1) are evaluated against external task success criteria, not against self-fitted parameters. Citations to Diffusion Policy [1], Torne et al. [6], Mark et al. [9], and related works supply baselines and prior claims being re-tested; none function as a uniqueness theorem or ansatz that forces the reported rankings by construction. Minor self-citations (e.g., Wei et al. [18] for push-T data) are ordinary methodological reuse and are not load-bearing for the corrective claim against prior literature. Checkpoint selection and single-run variance are methodological limitations (correctness risk), not circularity. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 2 invented entities

Empirical robotics paper; load-bearing choices are architectural and curriculum hyperparameters plus the representativeness of the chosen task suite and data scales. No new physical entities; free parameters are training knobs.

free parameters (4)
  • N (task-specific demonstration count)
    Chosen per task to match 'usual' single-task data; all scaling claims (N/2, N, 2N) are relative to this hand-chosen baseline.
  • p=0.8 in Random Sprinkle curriculum
    Probability mass on the longest context; fixed by authors without sweep reported as primary.
  • short vs full past-prediction horizon Tpast_p
    Switched depending on data regime; choice affects reported gains of variable-history training.
  • UNet channel widths and DiT layer count
    DiT enlarged and causal attention removed to match parameter count; architecture comparison depends on these retunings.
axioms (3)
  • domain assumption Diffusion Policy denoising loss and action-chunk formulation of Chi et al. are the correct base learner.
    All experiments build on this; no comparison to other BC or RL backbones.
  • ad hoc to paper Local stability of the manipulation primitive (prehensile grasp vs planar push) primarily dictates sample complexity of long-context learning.
    Used to interpret why grasp-and-return succeeds at N/2 while push-and-return needs more data (Sec 3.2, D.2).
  • ad hoc to paper Best-of-checkpoints by closed-loop success is a fair estimator of policy quality.
    Stated in B.3; introduces selection bias relative to fixed-step or validation-only selection.
invented entities (2)
  • Variable History Training (Algorithm 1) no independent evidence
    purpose: Jointly train one policy over a curriculum of context lengths to reduce overfitting in low data.
    Core algorithmic contribution; no independent prior evidence outside this paper.
  • push-and-return / grasp-and-return tasks no independent evidence
    purpose: Controlled benchmarks that separate memory requirement from manipulation stability.
    New environments introduced for the study; success metrics (contextual success) defined here.

pith-pipeline@v1.1.0-grok45 · 24381 in / 2560 out tokens · 30545 ms · 2026-07-12T13:48:26.958606+00:00 · methodology

0 comments
read the original abstract

Imitation learning has enabled highly-dexterous robotic manipulation from RGB observations. Policies trained with these methods, however, typically condition robot actions on only a short history of observations. These policies cannot solve tasks that require memory and can get stuck repeatedly executing the same failing motions. In this work, we first benchmark policy performance as context length is incrementally increased from short to long, across a spectrum of tasks with varying local stability and memory requirements, and in multiple data regimes. To our knowledge, this is the first study to investigate context length for Diffusion Policies at this level of detail. Our results challenge prior claims: naively scaling context length is not as brittle as advertised in literature. With an appropriate conditioning method and denoising backbone (UNet+Cross-Attention), single-task policies achieve high success rates on many tasks in the usual data regime even with naive scaling. Next, we propose a training algorithm to jointly train policies at multiple context lengths, further reducing the sample complexity of long-context learning. Finally, we apply our findings to re-evaluate some previously proposed solutions to long-context imitation learning.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 1 canonical work pages

  1. [1]

    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, 0(0):02783649241273668, 0. doi:10.1177/02783649241273668. URLhttps:// doi.org/10.1177/02783649241273668

    doi:10.1177/02783649241273668

  2. [2]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware, 2023. URLhttps://arxiv.org/abs/2304.13705

    Pith/arXiv arXiv 2023

  3. [3]

    A careful examination of large behavior models for multitask dexterous manipulation, 2025

    TRI LBM Team. A careful examination of large behavior models for multitask dexterous manipulation, 2025. URLhttps://arxiv.org/abs/2507.05331

    Pith/arXiv arXiv 2025

  4. [4]

    Physical Intelligence.π 0.5: a vision-language-action model with open-world generalization,

  5. [5]

    URLhttps://arxiv.org/abs/2504.16054

    Pith/arXiv arXiv

  6. [6]

    Gemini: A family of highly capable multimodal models, 2024

    Gemini Team. Gemini: A family of highly capable multimodal models, 2024. URLhttps: //arxiv.org/abs/2312.11805

    Pith/arXiv arXiv 2024

  7. [7]

    Torne, A

    M. Torne, A. Tang, Y . Liu, and C. Finn. Learning long-context diffusion policies via past-token prediction, 2025. URLhttps://arxiv.org/abs/2505.09561

    Pith/arXiv arXiv 2025

  8. [8]

    Sridhar, J

    A. Sridhar, J. Pan, S. Sharma, and C. Finn. MemER: Scaling up memory for robot control via experience retrieval, 2025. URLhttps://arxiv.org/abs/2510.20328

    arXiv 2025

  9. [9]

    Torne, K

    M. Torne, K. Pertsch, H. Walke, K. Vedder, S. Nair, B. Ichter, A. Z. Ren, H. Wang, J. Tang, K. Stachowicz, K. Dhabalia, M. Equi, Q. Vuong, J. T. Springenberg, S. Levine, C. Finn, and D. Driess. MEM: Multi-scale embodied memory for vision language action models, 2026. URLhttps://arxiv.org/abs/2603.03596

    arXiv 2026

  10. [10]

    M. S. Mark, J. Liang, M. Attarian, C. Fu, D. Dwibedi, D. Shah, and A. Kumar. BPP: Long- context robot imitation learning by focusing on key history frames, 2026. URLhttps:// arxiv.org/abs/2602.15010

    arXiv 2026

  11. [11]

    Zheng, Y

    R. Zheng, Y . Liang, S. Huang, J. Gao, H. Daum ´e III, A. Kolobov, F. Huang, and J. Yang. TraceVLA: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies, 2025. URLhttps://arxiv.org/abs/2412.10345

    Pith/arXiv arXiv 2025

  12. [12]

    H. Shi, B. Xie, Y . Liu, L. Sun, F. Liu, T. Wang, E. Zhou, H. Fan, X. Zhang, and G. Huang. MemoryVLA: Perceptual-cognitive memory in vision-language-action models for robotic ma- nipulation, 2026. URLhttps://arxiv.org/abs/2508.19236

    Pith/arXiv arXiv 2026

  13. [13]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn. OpenVLA: An open-source vision-language-action model, 2024. URL https://arxiv.org/abs/2406.09246. 10

    Pith/arXiv arXiv 2024

  14. [14]

    Guhur, S

    P.-L. Guhur, S. Chen, R. Garcia, M. Tapaswi, I. Laptev, and C. Schmid. Instruction-driven history-aware policies for robotic manipulations, 2022. URLhttps://arxiv.org/abs/ 2209.04899

    Pith/arXiv arXiv 2022

  15. [15]

    de Haan, D

    P. de Haan, D. Jayaraman, and S. Levine. Causal confusion in imitation learning, 2019. URL https://arxiv.org/abs/1905.11979

    Pith/arXiv arXiv 2019

  16. [16]

    C. Wen, J. Lin, T. Darrell, D. Jayaraman, and Y . Gao. Fighting copycat agents in behavioral cloning from observation histories, 2020. URLhttps://arxiv.org/abs/2010.14876

    Pith/arXiv arXiv 2020

  17. [17]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models, 2020. URLhttps: //arxiv.org/abs/2006.11239

    Pith/arXiv arXiv 2020

  18. [18]

    Cadene, S

    R. Cadene, S. Alibert, A. Soare, Q. Gallouedec, A. Zouitine, S. Palma, P. Kooijmans, M. Ar- actingi, M. Shukor, D. Aubakirova, M. Russi, F. Capuano, C. Pascal, J. Choghari, J. Moss, and T. Wolf. LeRobot: State-of-the-art machine learning for real-world robotics in PyTorch. https://github.com/huggingface/lerobot, 2024

    2024

  19. [19]

    A. Wei, A. Agarwal, B. Chen, R. Bosworth, N. Pfaff, and R. Tedrake. Empirical analysis of sim-and-real cotraining of diffusion policies for planar pushing from pixels, 2025. URL https://arxiv.org/abs/2503.22634

    Pith/arXiv arXiv 2025

  20. [20]

    Mandlekar, D

    A. Mandlekar, D. Xu, J. Wong, S. Nasiriany, C. Wang, R. Kulkarni, L. Fei-Fei, S. Savarese, Y . Zhu, and R. Mart´ın-Mart´ın. What matters in learning from offline human demonstrations for robot manipulation. InConference on Robot Learning (CoRL), 2021

    2021

  21. [21]

    A. Z. Ren, J. Lidard, L. L. Ankile, A. Simeonov, P. Agrawal, A. Majumdar, B. Burchfiel, H. Dai, and M. Simchowitz. Diffusion policy policy optimization, 2024. URLhttps:// arxiv.org/abs/2409.00588

    Pith/arXiv arXiv 2024

  22. [22]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation, 2015. URLhttps://arxiv.org/abs/1505.04597

    Pith/arXiv arXiv 2015

  23. [23]

    Perez, F

    E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. Courville. FiLM: Visual reasoning with a general conditioning layer, 2017. URLhttps://arxiv.org/abs/1709.07871

    Pith/arXiv arXiv 2017

  24. [24]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image syn- thesis with latent diffusion models, 2022. URLhttps://arxiv.org/abs/2112.10752

    Pith/arXiv arXiv 2022

  25. [25]

    T. J. Boerner, S. Deems, T. R. Furlani, S. L. Knuth, and J. Towns. Access: Advancing innovation: Nsf’s advanced cyberinfrastructure coordination ecosystem: Services & sup- port. InPractice and Experience in Advanced Research Computing 2023: Computing for the Common Good, PEARC ’23, page 173–176, New York, NY , USA, 2023. Association for Computing Machiner...

    doi:10.1145/3569951.3597559 2023

  26. [26]

    Reuther, J

    A. Reuther, J. Kepner, C. Byun, S. Samsi, W. Arcand, D. Bestor, B. Bergeron, V . Gadepally, M. Houle, M. Hubbell, M. Jones, A. Klein, L. Milechin, J. Mullen, A. Prout, A. Rosa, C. Yee, and P. Michaleas. Interactive supercomputing on 40,000 cores for machine learning and data analysis. In2018 IEEE High Performance extreme Computing Conference (HPEC), page ...

    2018

  27. [27]

    K. Song, B. Chen, M. Simchowitz, Y . Du, R. Tedrake, and V . Sitzmann. History-guided video diffusion, 2025. URLhttps://arxiv.org/abs/2502.06764

    Pith/arXiv arXiv 2025

  28. [28]

    Z. Li, X. B. Peng, P. Abbeel, S. Levine, G. Berseth, and K. Sreenath. Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control, 2024. URLhttps://arxiv. org/abs/2401.16889. 11

    Pith/arXiv arXiv 2024

  29. [29]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, and J. Malik. RMA: Rapid motor adaptation for legged robots,

  30. [30]

    URLhttps://arxiv.org/abs/2107.04034

    Pith/arXiv arXiv

  31. [31]

    X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic control with dynamics randomization. In2018 IEEE International Conference on Robotics and Automation (ICRA), page 3803–3810. IEEE, May 2018. doi:10.1109/icra.2018.8460528. URLhttp://dx.doi.org/10.1109/ICRA.2018.8460528

    doi:10.1109/icra.2018.8460528 2018

  32. [32]

    R. P. Singh, Z. Xie, P. Gergondet, and F. Kanehiro. Learning bipedal walking for humanoids with current feedback.IEEE Access, 11:82013–82023, 2023. ISSN 2169-3536. doi:10.1109/ access.2023.3301175. URLhttp://dx.doi.org/10.1109/ACCESS.2023.3301175

    work page doi:10.1109/access.2023.3301175 2023

  33. [33]

    N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, and P. Liang. Lost in the middle: How language models use long contexts, 2023. URLhttps://arxiv.org/abs/ 2307.03172

    Pith/arXiv arXiv 2023

  34. [34]

    J. Li, M. Wang, Z. Zheng, and M. Zhang. Loogle: Can long-context language models under- stand long contexts?, 2024. URLhttps://arxiv.org/abs/2311.04939

    Pith/arXiv arXiv 2024

  35. [35]

    G. Qin, Y . Feng, and B. Van Durme. The NLP task effectiveness of long-range transformers. InProceedings of the 17th Conference of the European Chapter of the Association for Compu- tational Linguistics. Association for Computational Linguistics, 2023. doi:10.18653/v1/2023. eacl-main.273. URLhttp://dx.doi.org/10.18653/v1/2023.eacl-main.273

    doi:10.18653/v1/2023 2023

  36. [36]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition, 2015. URL https://arxiv.org/abs/1512.03385

    Pith/arXiv arXiv 2015

  37. [37]

    Tedrake and the Drake Development Team

    R. Tedrake and the Drake Development Team. Drake: Model-based design and verification for robotics, 2019. URLhttps://drake.mit.edu

    2019

  38. [38]

    Tedrake.Robotic Manipulation

    R. Tedrake.Robotic Manipulation. 2024. URLhttp://manipulation.mit.edu

    2024

  39. [39]

    S. Ross, G. J. Gordon, and J. A. Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning, 2011. URLhttps://arxiv.org/abs/1011.0686. A Additional Related Works A.1 Context Length in Imitation Learning We have already introduced and discussed the work from Torne et al. [6], Mark et al. [9], Sridhar et al. [7]. We no...

    Pith/arXiv arXiv 2011

  40. [40]

    provide ablations on comparing feed-forward policies with history with LSTM based policies

  41. [41]

    A.3 Context Length Understanding in LLMs We believe the robotics community needs to build an understanding of policy learning in the pres- ence of longer contexts

    argue that a non-state history based LSTM better generalizes to dynamics of a physical system, while [30] choose a non-history based policy as the best performer based on their ablations. A.3 Context Length Understanding in LLMs We believe the robotics community needs to build an understanding of policy learning in the pres- ence of longer contexts. LLM c...

  42. [42]

    This, we believe, would also give more consistent results across architectures, checkpoints, and context lengths and is a very relevant direction for future work. D.3.2 Manipulation Skill Learning In Section 3, we note that in the low data regime,grasp-and-returnprovides good long-context per- formance, where aspush-and-returnperformance drops. We now com...

  43. [43]

    but rather present in the original work [1], we also investigate if predicting past actions helps with short-context policy learning. 22 (a) (b) (c) (d) (e) (f) Figure 15: Comparison ofUNet+Cross-Attention,UNet+FiLM, andvariable history training method introduced above across tasks (data scaleN/2in left column andNin right column).Vari- able historymethod...