Pith. sign in

REVIEW 2 major objections 3 minor 69 references

ChainVLA claims that long-horizon manipulation succeeds only when each new prediction is conditioned on a joint, revisable state carrying both task progress and the previous prediction's unfinished motion.

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 · deepseek-v4-flash

2026-08-04 09:15 UTC pith:JNDUOKQZ

load-bearing objection A genuine architectural combination with matched ablations and honest limitations, but the exact margins rest on a single seed and non-rerun baselines — worth a serious referee, not yet canonical. the 2 major comments →

arxiv 2608.02326 v1 pith:JNDUOKQZ submitted 2026-08-03 cs.RO

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

Yuzhi Huang , Weijue Bu , Ziyi Xiong , Jie Wu , Fanding Huang , Jingyan Jiang , Zhi Wang This is my paper
classification cs.RO
keywords vision-language-action policylong-horizon manipulationaction chunkingcross-query execution statemotion continuationsparse event memoryrecurrent working statereceding-horizon control
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.

This paper argues that action-chunked vision-language-action (VLA) policies fail at long-horizon manipulation because each replanning query throws away two things: task-relevant evidence from earlier observations and the unexecuted remainder of the previous action prediction. ChainVLA carries both across queries in one execution state—Progress Context, built from a recurrent working state and sparse event memory, plus Motion Tail, the preceding prediction's unexecuted suffix—and regenerates each action horizon under the latest observation. On memory-dependent RMBench tasks it reports 62.8% average success, while removing Motion Tail drops success to 11.2% and removing Progress Context to 3.0%. The asymmetric drop is presented as evidence that motion continuity preserves the observation stream from which task progress is inferred. If the mechanism holds, any receding-horizon policy should be chained through a comparable state rather than replanning from the current input alone.

Core claim

In the strict-prefix receding-horizon regime, where a policy predicts a finite action sequence, executes a short prefix, and replans, the current input can be decision-insufficient: identical observations and instructions can follow different histories and require different next actions. ChainVLA's central claim is that the missing quantity is a policy-side execution state spanning queries, composed of a retrospective half (Progress Context) carrying observation-derived task progress and a prospective half (Motion Tail) carrying the previous prediction's unexecuted suffix. Each query reads, updates, and writes this state, while a conditional-flow decoder regenerates every horizon position un

What carries the argument

The load-bearing object is the execution state s_k = (g_k, u_k) and its query transition. g_k is Progress Context, formed by fusing live tokens from a recurrent Working State with evidence retrieved from a sparse event memory; u_k is Motion Tail, the unexecuted suffix of the previous prediction. The suffix enters in two ways: encoded tokens condition Working State construction, and an aligned trajectory initializes the decoder's full-horizon state with additive noise. Because the decoder regenerates every position, the carried state guides rather than pins the next prediction; after decoding, a linear blend of overlapping steps smooths the executed handoff. This combination—reading, updating

Load-bearing premise

The load-bearing premise is that one random seed and 100 rollout episodes per RMBench task yield stable success-rate estimates, and that external baseline rows not rerun under the shared protocol are comparable enough for the headline margin to be meaningful.

What would settle it

Re-run the five RMBench tasks and the two main ablations across at least ten seeds and report the full distribution; if the 62.8 vs 11.2/3.0 gaps overlap across seeds, the chaining claim is not established. A cleaner mechanistic test: keep memory fixed and feed a precomputed smooth continuation as Motion Tail; if success stays near zero, conditioning the decoder before generation—not seam smoothness—is the active ingredient.

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

If this is right

  • Receding-horizon VLA policies are leaving recoverable history on the table; chaining a joint execution state can recover it without a separate planner.
  • Neither task memory alone nor motion continuation alone is sufficient on memory-dependent tasks; the complete two-part state is the load-bearing mechanism.
  • Post-decoding smoothing or ensembling cannot replace carrying the suffix into state construction and generation, because smoothing acts only on already-decoded commands.
  • The asymmetry between the two ablations is consistent with motion continuity protecting the observation stream that task-progress inference reads.
  • The state update is a single per-query transition shared by training and deployment, so the approach stays within the standard horizon-based policy family.

Where Pith is reading between the lines

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

  • If the mechanism generalizes, any policy that replans from scratch—not just manipulation VLAs—could benefit from conditioning on the previous prediction's unexecuted continuation.
  • The motion-continuity-protects-memory hypothesis predicts that deliberately perturbing the viewpoint trajectory of a memory-only policy should reproduce much of the 62.8-to-11.2 collapse; that experiment would test the claimed causal ordering.
  • The single-seed, fixed-protocol evaluation means the headline margins should be treated as point estimates; re-estimating them across seeds is the direct check on the central claim.
  • A natural extension is latency-robust and hardware deployment, where the boundary diagnostics would measure commanded rather than physical trajectories; the paper's fixed-rate simulation leaves that untested.

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

2 major / 3 minor

Summary. The paper proposes ChainVLA, a 1.2B-parameter VLA policy that maintains a joint, revisable execution state across receding-horizon queries. The state has two components: Progress Context (a recurrent Working State plus a Sparse Event Memory carrying observation-derived task progress) and Motion Tail (the unexecuted suffix of the preceding prediction, fed into state construction and decoder initialization). The authors report 62.8% average success on RMBench and 98.8% across LIBERO, with ablations dropping RMBench to 11.2% when Motion Tail is removed, 3.0% when Progress Context is removed, and 1.6% when both are removed. They interpret these asymmetric ablations as evidence that the joint state, not either component alone, is load-bearing, with motion continuity helping preserve the observation stream used for task-progress inference.

Significance. If the point estimates are reliable, the paper makes a useful architectural contribution: it unifies long-term task memory and short-term motion continuation in a single per-query transition, rather than treating them as separate add-ons. The internal evaluation is more careful than many VLA papers: matched ablations are specified precisely (Table A3), post-decoding controls (Linear Continuation, Temporal Ensemble) isolate model-side conditioning from mere smoothing, and the boundary diagnostics are clearly defined. The paper also explicitly lists its limitations in Appendix D. The main weakness is statistical: the central ablation claim rests on single-seed point estimates, and the headline comparison to Mem-0 uses source-reported numbers that were not rerun under the shared protocol.

major comments (2)
  1. [Appendix D / Table A2; §4.3/Table 2] The central claim that removing Motion Tail or Progress Context collapses RMBench success (62.8% to 11.2% and 3.0%) rests entirely on point estimates from a single random seed (seed 42, shared by all internally trained conditions). Appendix D explicitly states that no seed variability or significance tests are reported. Because every condition shares the same seed, the non-additivity that motivates the 'joint state' interpretation could be inflated by seed-specific training dynamics; a different seed might compress or enlarge the gaps. Please rerun at least Full, w/o Motion Tail, and w/o Progress Context with multiple seeds (and ideally w/o Both) and report means with confidence intervals or significance tests. If multiple seeds are infeasible, the abstract and main text should be rephrased to make clear these are single-seed point estimates and avoid language implying a robust quantitat
  2. [§4.2/Table 1 and Appendix D] The headline comparison to Mem-0 (62.8 vs. 52.8) is not protocol-matched. Appendix D states that the source-reported comparison rows were not rerun under the authors' protocol and may differ in backbone, demonstration count, optimization budget, and success criterion. The main text nevertheless presents ChainVLA as 'ahead of the strongest listed method with complete five-task coverage.' This is a positioning statement, not an established result under matched conditions. Either rerun the strongest external baselines (at least Mem-0) under the shared protocol, or explicitly restrict all comparative claims to the internally matched ablation table and describe the external rows as non-comparable references. As written, the claimed 10-point margin over Mem-0 is not supported.
minor comments (3)
  1. [Throughout] Several passages contain missing spaces or typos, likely from PDF extraction: e.g., the abstract has 'retainingknowledgeofwhatearlieractionshaveestablished', and §3.4 has 'Ialigns'. These should be corrected in the camera-ready version.
  2. [Table 1] The MemoAct row appears garbled: 'MemoAct 49841 –55–'. The intended per-task values are unclear. Please reformat the table so each row has the same structure.
  3. [§4.2] The text says 'ahead of the strongest listed method' but the external rows are not rerun; consider adding a footnote or pointer to Appendix D at the first mention of Mem-0 so the reader immediately understands the comparability caveat.

Circularity Check

0 steps flagged

No derivation-level circularity; central claims are tested against external benchmarks and matched ablations, with only a non-load-bearing self-citation.

full rationale

The paper's central claims are empirical: ChainVLA's success is measured on external benchmarks (RMBench, LIBERO) and its ablations are matched interventions that withhold specific inputs, training terms, and deployment handling together. Nothing in the method section derives a prediction from a fitted quantity: Motion Tail is literally the previous prediction's unexecuted suffix (Eq. 3, uk+1 = Tail(Ak)), and conditioning the next horizon on it is a recurrent state, not a definitional equivalence between input and output. Progress Context is similarly built from observations and retrieved memory, and the stage labels supervise event writes but do not enter the action decoder. The one self-citation, to the authors' own RoboStream among a list of memory-augmented policies in Related Work, is not load-bearing: it is used only to position prior work, not to justify ChainVLA's design or results. No uniqueness theorem, ansatz, or known result is imported from same-author work. The limitations stated in Appendix D — single random seed, no significance tests, and source-reported baseline rows not rerun under the shared protocol — are statistical and comparability concerns, not circularity. They affect how much confidence the point estimates deserve but do not make the central 62.8% vs 11.2%/3.0% contrast equivalent to the method's own assumptions by construction. The paper is transparent that boundary metrics establish association rather than causation, and that external rows are used for positioning only. Thus there is no circular step to exhibit: the derivation chain is self-contained against external evaluation. Score 1 reflects the minor same-author citation in related work, which is not load-bearing and does not raise the substantive circularity score to 2 or above.

Axiom & Free-Parameter Ledger

7 free parameters · 7 axioms · 4 invented entities

The central claim rests on benchmark validity, architecture expressiveness, and an untested single-seed protocol. The method introduces several hand-fixed hyperparameters and latent state components, each with only internal ablations as evidence.

free parameters (7)
  • Horizon H and executed prefix h_exec = 30 / 24
    Fixed once; defines the size of the unexecuted Motion Tail and the replanning frequency. No sensitivity reported.
  • Sparse Event Memory capacity = 16 records
    Fixed design choice; oldest non-anchor evicted. No sensitivity reported.
  • Write-confidence threshold = 0.55
    Stage-triggered rollout writes require predicted stage confidence at least 0.55.
  • Periodic write interval = 3 queries
    Used for annotation-free data and for w/o Stage Annotation; fixed and reused across conditions.
  • Overlap-consistency loss weight = 0.2
    Regularizer for Motion Tail handoff during training; fixed once.
  • Tail application probability during training = 0.5
    Each eligible query uses its predecessor's detached predicted tail with probability 0.5, otherwise the tail path is masked.
  • Motion Tail noise scale sigma_u = not reported
    Sets the additive noise for tail-aligned action initialization; value is not specified in the paper.
axioms (7)
  • domain assumption Current input plus carried execution state is sufficient to approximate the optimal history-conditioned policy (Eq. 2).
    The whole architecture is built around this approximation; no proof of sufficiency is given.
  • domain assumption RMBench's decision-insufficiency criterion (Eq. 1) correctly identifies tasks where the current observation is not enough for the optimal next action.
    The paper relies on the external benchmark to make the memory problem real.
  • domain assumption RMBench stage annotations supervise event writes without leaking the task solution into action generation.
    The paper states stage labels never enter the action decoder, but stage supervision still shapes the memory pathway.
  • ad hoc to paper One random seed is representative of training and evaluation outcome.
    Appendix D reports no seed variability or significance tests; all internally trained conditions use a single seed.
  • ad hoc to paper Fixed design choices (capacity 16, threshold 0.55, interval 3, overlap weight 0.2) are adequate and do not require sensitivity analysis.
    The paper fixed these once and reused them; no sensitivity to them is reported.
  • standard math Standard conditional-flow / diffusion-transformer training is a valid action-generation objective.
    The action decoder is trained with the standard conditional-flow action objective.
  • domain assumption External baseline rows from cited papers are comparable enough for positioning claims.
    The paper explicitly says these rows were not rerun under its protocol and may differ in backbone, demos, optimization budget, and success criterion.
invented entities (4)
  • Execution state s_k = (g_k, u_k) no independent evidence
    purpose: Carry task progress and unfinished motion across receding-horizon queries in one read-update-write transition.
    Defined by this paper; no external falsifiable prediction beyond the internal ablations.
  • Progress Context g_k no independent evidence
    purpose: Fuse live working-state tokens with retrieved event evidence to represent observation-derived task progress.
    Internal latent whose effectiveness is shown only by within-paper ablations.
  • Motion Tail u_k no independent evidence
    purpose: Carry the preceding prediction's unexecuted suffix into state construction and action generation.
    It is a directly defined object (the unexecuted suffix of the previous prediction), but its effectiveness is only shown by internal ablations.
  • Sparse Event Memory C_k with event records no independent evidence
    purpose: Store sparse task milestones for later retrieval, with age-aware scoring and trigger-based writes.
    Memory concept is common in the literature, but this paper's specific write/retrieval design has no external validation.

pith-pipeline@v1.3.0-daily-deepseek · 15953 in / 16303 out tokens · 134326 ms · 2026-08-04T09:15:12.034372+00:00 · methodology

0 comments
read the original abstract

Humans perform long-horizon manipulation by retaining knowledge of what earlier actions have established while continuously adapting the motion underway. By contrast, action-chunked vision-language-action (VLA) policies repeatedly replan from the current input at each query. Existing methods preserve either long-term task evidence through memory or short-term motion through action reuse and ensembling, leaving the cross-query handoff incomplete. We introduce ChainVLA, a 1.2B-parameter VLA policy that chains successive queries through a joint and revisable execution state. Progress Context combines a recurrent Working State with sparse event memory to carry observation-derived task progress, while Motion Tail feeds the preceding prediction's unexecuted continuation into state construction and action generation. Together, the two components condition a decoder that regenerates each action horizon under the latest observation, allowing the carried state to guide the next prediction without fixing it. ChainVLA reaches 62.8% average success on RMBench and 98.8% across four LIBERO suites, while removing Motion Tail or Progress Context reduces RMBench success to 11.2% and 3.0%, respectively. These asymmetric ablations are consistent with motion continuity helping preserve the observation stream from which task progress is inferred.

Figures

Figures reproduced from arXiv: 2608.02326 by Fanding Huang, Jie Wu, Jingyan Jiang, Weijue Bu, Yuzhi Huang, Zhi Wang, Ziyi Xiong.

Figure 1
Figure 1. Figure 1: Motivation and design of ChainVLA. (a) At each query boundary, earlier task evidence may be absent from the current [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall query-transition architecture of ChainVLA. At query [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustrative Put Back Block boundary from paired [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Boundary errors for Full and w/o Motion Tail. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

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

69 extracted references · 1 canonical work pages

  1. [1]

    str \"o m, K. J. 1965. Optimal Control of Markov Processes with Incomplete State Information. Journal of Mathematical Analysis and Applications, 10: 174--205

    1965

  2. [2]

    Y.; and Levine, S

    Black, K.; Galliker, M. Y.; and Levine, S. 2025. Real-Time Execution of Action Chunking Flow Policies. arXiv:2506.07339

    Pith/arXiv arXiv 2025

  3. [3]

    Black, K.; et al. 2025. _0 : A Vision-Language-Action Flow Model for General Robot Control. In Proceedings of Robotics: Science and Systems

    2025

  4. [5]

    Chen, Y.; et al. 2026 b . Non-Markovian Long-Horizon Robot Manipulation via Keyframe Chaining. arXiv:2603.01465

    arXiv 2026

  5. [6]

    Chi, C.; Xu, Z.; Feng, S.; Cousineau, E.; Du, Y.; Burchfiel, B.; Tedrake, R.; and Song, S. 2025. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. The International Journal of Robotics Research, 44(10--11): 1684--1704

    2025

  6. [7]

    Chung, N.; et al. 2026. Rethinking Progression of Memory State in Robotic Manipulation: An Object-Centric Perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, 3407--3415

    2026

  7. [8]

    Dai, Y.; Fu, H.; Lee, J.; Liu, Y.; Zhang, H.; Yang, J.; Finn, C.; Fazeli, N.; and Chai, J. 2026. RoboMME : Benchmarking and Understanding Memory for Robotic Generalist Policies. In Proceedings of the International Conference on Machine Learning. Oral presentation

    2026

  8. [9]

    Guo, X.; et al. 2026. Chameleon: Control-Indexed Prospective Memory for Visuomotor Manipulation. arXiv:2603.24576

    Pith/arXiv arXiv 2026

  9. [10]

    Hu, Q.; Qiu, Z.; Xu, Z.; Zhang, K.; Bu, X.; Sun, Z.; Zhang, B.; Zhao, J.; Gan, Z.; and Ding, W. 2026. Resolving State Ambiguity in Robot Manipulation via Adaptive Working Memory Recoding. IEEE Robotics and Automation Letters, 11(7): 8439--8446

    2026

  10. [12]

    P.; Littman, M

    Kaelbling, L. P.; Littman, M. L.; and Cassandra, A. R. 1998. Planning and Acting in Partially Observable Stochastic Domains. Artificial Intelligence, 101(1--2): 99--134

    1998

  11. [13]

    J.; et al

    Kim, M. J.; et al. 2024. OpenVLA : An Open-Source Vision-Language-Action Model. In 8th Annual Conference on Robot Learning

    2024

  12. [14]

    Lashley, K. S. 1951. The Problem of Serial Order in Behavior. In Jeffress, L. A., ed., Cerebral Mechanisms in Behavior: The Hixon Symposium, 112--136. New York: Wiley

    1951

  13. [18]

    Liu, B.; Zhu, Y.; Gao, C.; Feng, Y.; Liu, Q.; Zhu, Y.; and Stone, P. 2023. LIBERO : Benchmarking Knowledge Transfer for Lifelong Robot Learning. In Advances in Neural Information Processing Systems

    2023

  14. [19]

    Liu, Y.; Ibn Hamid, J.; Xie, A.; Lee, Y.; Du, M.; and Finn, C. 2024. Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling. arXiv:2408.17355

    Pith/arXiv arXiv 2024

  15. [20]

    Liu, Y.; Yu, H.; Zhao, J.; Li, B.; Zhang, D.; Li, M.; Wu, W.; Hu, Y.; Xie, J.; Guo, J.; Wang, D.; and Gao, Y. 2026 a . Learning Native Continuation for Action Chunking Flow Policies. arXiv:2602.12978

    Pith/arXiv arXiv 2026

  16. [22]

    Malhotra, R.; Liu, Y.; and Finn, C. 2025. Self-Guided Action Diffusion. arXiv:2508.12189

    Pith/arXiv arXiv 2025

  17. [23]

    A.; Galanter, E.; and Pribram, K

    Miller, G. A.; Galanter, E.; and Pribram, K. H. 1960. Plans and the Structure of Behavior. New York: Holt, Rinehart and Winston

    1960

  18. [24]

    NVIDIA ; Bjorck, J.; Castaneda, F.; Cherniadev, N.; Da, X.; Ding, R.; Fan, L.; et al. 2025. GR00T N1 : An Open Foundation Model for Generalist Humanoid Robots. arXiv:2503.14734

    Pith/arXiv arXiv 2025

  19. [26]

    Physical Intelligence ; Black, K.; Brown, N.; Darpinian, J.; Dhabalia, K.; Driess, D.; et al. 2025. _ 0.5 : A Vision-Language-Action Model with Open-World Generalization. In 9th Annual Conference on Robot Learning. Oral presentation

    2025

  20. [29]

    Shi, H.; et al. 2026 b . MemoryVLA : Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic Manipulation. In The Fourteenth International Conference on Learning Representations

    2026

  21. [30]

    D.; and Sondik, E

    Smallwood, R. D.; and Sondik, E. J. 1973. The Optimal Control of Partially Observable Markov Processes over a Finite Horizon. Operations Research, 21(5): 1071--1088

    1973

  22. [36]

    M.; Ghahramani, Z.; and Jordan, M

    Wolpert, D. M.; Ghahramani, Z.; and Jordan, M. I. 1995. An Internal Model for Sensorimotor Integration. Science, 269(5232): 1880--1882

    1995

  23. [37]

    Xiao, B.; Wu, H.; Xu, W.; Dai, X.; Hu, H.; Lu, Y.; Zeng, M.; Liu, C.; and Yuan, L. 2024. Florence-2 : Advancing a Unified Representation for a Variety of Vision Tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4818--4829

    2024

  24. [42]

    Z.; Kumar, V.; Levine, S.; and Finn, C

    Zhao, T. Z.; Kumar, V.; Levine, S.; and Finn, C. 2023. Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. In Proceedings of Robotics: Science and Systems XIX

    2023

  25. [43]

    Zheng, J.; et al. 2026. X-VLA : Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model. In The Fourteenth International Conference on Learning Representations

    2026

  26. [44]

    Zitkovich, Brianna and others , booktitle =

  27. [45]

    Kim, Moo Jin and others , booktitle =

  28. [46]

    doi:10.1109/CVPR52733.2024.00461 , year =

    Xiao, Bin and Wu, Haiping and Xu, Weijian and Dai, Xiyang and Hu, Houdong and Lu, Yumao and Zeng, Michael and Liu, Ce and Yuan, Lu , booktitle =. doi:10.1109/CVPR52733.2024.00461 , year =

    arXiv 2024

  29. [47]

    Black, Kevin and others , booktitle =

  30. [48]

    9th Annual Conference on Robot Learning , note =

  31. [49]

    The International Journal of Robotics Research , volume =

    Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author =. The International Journal of Robotics Research , volume =. doi:10.1177/02783649241273668 , year =

    doi:10.1177/02783649241273668

  32. [50]

    Proceedings of Robotics: Science and Systems XIX , doi =

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware , author =. Proceedings of Robotics: Science and Systems XIX , doi =

  33. [51]

    Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter , booktitle =

  34. [52]

    Zheng, Jinliang and others , booktitle =

  35. [53]

    2501.15830 , archivePrefix =

    Qu, Delin and Song, Haoming and Chen, Qizhi and Yao, Yuanqi and Ye, Xinyi and Ding, Yan and Wang, Zhigang and Gu, JiaYuan and Zhao, Bin and Wang, Dong and Li, Xuelong , year =. 2501.15830 , archivePrefix =

    Pith/arXiv arXiv

  36. [54]

    2501.09747 , archivePrefix =

    Pertsch, Karl and Stachowicz, Kyle and Ichter, Brian and Driess, Danny and Nair, Suraj and Vuong, Quan and Mees, Oier and Finn, Chelsea and Levine, Sergey , year =. 2501.09747 , archivePrefix =

    Pith/arXiv arXiv

  37. [55]

    Journal of Mathematical Analysis and Applications , volume =

    Optimal Control of Markov Processes with Incomplete State Information , author =. Journal of Mathematical Analysis and Applications , volume =. doi:10.1016/0022-247X(65)90154-X , year =

    doi:10.1016/0022-247x(65)90154-x

  38. [56]

    Operations Research , volume =

    The Optimal Control of Partially Observable Markov Processes over a Finite Horizon , author =. Operations Research , volume =. doi:10.1287/opre.21.5.1071 , year =

    doi:10.1287/opre.21.5.1071

  39. [57]

    Artificial Intelligence , volume =

    Planning and Acting in Partially Observable Stochastic Domains , author =. Artificial Intelligence , volume =. doi:10.1016/S0004-3702(98)00023-X , year =

    doi:10.1016/s0004-3702(98)00023-x

  40. [58]

    Plans and the Structure of Behavior , author =

  41. [59]

    Consciousness and Self-Regulation: Advances in Research and Theory , editor =

    Attention to Action: Willed and Automatic Control of Behavior , author =. Consciousness and Self-Regulation: Advances in Research and Theory , editor =

  42. [60]

    Shi, Hao and others , booktitle =

  43. [61]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Rethinking Progression of Memory State in Robotic Manipulation: An Object-Centric Perspective , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. doi:10.1609/aaai.v40i5.37337 , year =

    work page doi:10.1609/aaai.v40i5.37337

  44. [62]

    Dai, Yinpei and Fu, Hongze and Lee, Jayjun and Liu, Yuejiang and Zhang, Haoran and Yang, Jianing and Finn, Chelsea and Fazeli, Nima and Chai, Joyce , booktitle =

  45. [63]

    2603.01229 , archivePrefix =

    Chen, Tianxing and Wang, Yuran and Li, Mingleyang and Qin, Yan and Shi, Hao and Li, Zixuan and Hu, Yifan and Zhang, Yingsheng and Wang, Kaixuan and Chen, Yue and others , year =. 2603.01229 , archivePrefix =

    Pith/arXiv arXiv

  46. [64]

    2510.20328 , archivePrefix =

    Sridhar, Ajay and Pan, Jennifer and Sharma, Satvik and Finn, Chelsea , year =. 2510.20328 , archivePrefix =

    arXiv

  47. [65]

    2603.04910 , archivePrefix =

    Lei, Yuheng and others , year =. 2603.04910 , archivePrefix =

    Pith/arXiv arXiv

  48. [66]

    2026 , eprint =

    Non-Markovian Long-Horizon Robot Manipulation via Keyframe Chaining , author =. 2026 , eprint =

    2026

  49. [67]

    2603.07647 , archivePrefix =

    Sun, Jun and others , year =. 2603.07647 , archivePrefix =

    arXiv

  50. [68]

    2603.12942 , archivePrefix =

    Li, Hang and others , year =. 2603.12942 , archivePrefix =

    arXiv

  51. [69]

    2026 , eprint =

    Chameleon: Control-Indexed Prospective Memory for Visuomotor Manipulation , author =. 2026 , eprint =

    2026

  52. [70]

    2603.18494 , archivePrefix =

    Tan, Liufan and Li, Jiale and Jing, Gangshan , year =. 2603.18494 , archivePrefix =

    Pith/arXiv arXiv

  53. [71]

    IEEE Robotics and Automation Letters , volume =

    Resolving State Ambiguity in Robot Manipulation via Adaptive Working Memory Recoding , author =. IEEE Robotics and Automation Letters , volume =. doi:10.1109/LRA.2026.3693565 , year =

    arXiv 2026

  54. [72]

    2604.13942 , archivePrefix =

    Liu, Zhen and others , year =. 2604.13942 , archivePrefix =

    Pith/arXiv arXiv

  55. [73]

    Torne, Marcel and Pertsch, Karl and Walke, Homer and Vedder, Kyle and Nair, Suraj and Ichter, Brian and Ren, Allen Z. and Wang, Haohuan and Tang, Jiaming and Stachowicz, Kyle and Dhabalia, Karan and Equi, Michael and Vuong, Quan and Springenberg, Jost Tobias and Levine, Sergey and Finn, Chelsea and Driess, Danny , year =. 2603.03596 , archivePrefix =

    arXiv

  56. [74]

    2506.19816 , archivePrefix =

    Li, Hao and Yang, Shuai and Chen, Yilun and Chen, Xinyi and Yang, Xiaoda and Tian, Yang and Wang, Hanqing and Wang, Tai and Lin, Dahua and Zhao, Feng and Pang, Jiangmiao , year =. 2506.19816 , archivePrefix =

    arXiv

  57. [75]

    2604.18791 , archivePrefix =

    Zeng, Zijian and Ding, Fei and Yang, Huiming and Li, Xianwei , year =. 2604.18791 , archivePrefix =

    Pith/arXiv arXiv

  58. [76]

    2025 , eprint =

    Real-Time Execution of Action Chunking Flow Policies , author =. 2025 , eprint =

    2025

  59. [77]

    2026 , eprint =

    Learning Native Continuation for Action Chunking Flow Policies , author =. 2026 , eprint =

    2026

  60. [78]

    2024 , eprint =

    Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling , author =. 2024 , eprint =

    2024

  61. [79]

    2025 , eprint =

    Self-Guided Action Diffusion , author =. 2025 , eprint =

    2025

  62. [80]

    2606.09827 , archivePrefix =

    Shi, Hao and Li, Weiye and Xie, Bin and Wang, Yulin and Zhou, Renping and Wang, Tiancai and Zhang, Xiangyu and Luo, Ping and Huang, Gao , year =. 2606.09827 , archivePrefix =

    Pith/arXiv arXiv

  63. [81]

    2606.10363 , archivePrefix =

    Sun, Xiaoquan and Zhang, Ruijian and Cao, Chen and Sun, Yihan and Chen, Jiahui and Xu, Zetian and Chen, Bo and Chen, Haijier and Yang, Zhen and Zhu, Jiarun and Hong, Yijun and Xu, JingZhe and Pang, Jingrui and Yuan, Mingqi and Chen, Jiayu , year =. 2606.10363 , archivePrefix =

    Pith/arXiv arXiv

  64. [82]

    2606.20092 , archivePrefix =

    Yang, Ganlin and Tu, Zhangzheng and Yang, Yuqiang and Mao, Sitong and Dong, Junyi and Chen, Tianxing and Peng, Jiaqi and Xiong, Jing and Cao, Jiafei and Dai, Jifeng and Zhou, Wengang and Mu, Yao and Wang, Tai , year =. 2606.20092 , archivePrefix =

    Pith/arXiv arXiv

  65. [83]

    2606.23589 , archivePrefix =

    Zeng, Yihan and Ye, Minghao and Chen, Yiyuan and Shentu, Yide and Wu, Philipp and Yan, Zike and Li, Zhongyu , year =. 2606.23589 , archivePrefix =

    Pith/arXiv arXiv

  66. [84]

    2607.04609 , archivePrefix =

    Zhan, Dijia and Xu, Xuemiao and Li, Jinyi and Tang, Jie , year =. 2607.04609 , archivePrefix =

    Pith/arXiv arXiv

  67. [85]

    Cerebral Mechanisms in Behavior: The Hixon Symposium , editor =

    The Problem of Serial Order in Behavior , author =. Cerebral Mechanisms in Behavior: The Hixon Symposium , editor =

  68. [86]

    Science , volume =

    An Internal Model for Sensorimotor Integration , author =. Science , volume =

  69. [87]

    2603.12939 , archivePrefix =

    Huang, Yuzhi and Wu, Jie and Bu, Weijue and Xiong, Ziyi and Jiang, Gaoyang and Li, Ye and Ji, Kangye and Xie, Shuzhao and Huang, Yue and Wu, Chenglei and Jiang, Jingyan and Wang, Zhi , year =. 2603.12939 , archivePrefix =

    Pith/arXiv arXiv