Pith. sign in

REVIEW 3 major objections 5 minor 49 references

AC-VLA claims that vision-language-action robot models fail on out-of-distribution tasks because they memorize whole trajectories and rely on wrist-camera textures, and that decomposing demonstrations into sub-task segments while masking wr

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

2026-08-01 22:31 UTC pith:6WMHBLZQ

load-bearing objection Strong internal ablations and real-robot checks make the OOD claim plausible, but the paper overstates SOTA and leaves its core decomposition matching unspecified. the 3 major comments →

arxiv 2607.15714 v1 pith:6WMHBLZQ submitted 2026-07-17 cs.RO

AC-VLA: Robust Out-of-Distribution Action Execution via Compositional Learning

Xiaojiang Peng , Kai Peng , Jie Lu , Zheng Lian , Zitong YU , Xiaobo Wang This is my paper
classification cs.RO
keywords vision-language-action modelscompositional generalizationout-of-distributionrobotic manipulationtrajectory decompositionasymmetric maskingdense sub-task supervisionflow-matching policy
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.

AC-VLA is a training-time recipe that bolts onto an existing vision-language-action (VLA) policy without architectural changes. The paper's claim is that VLA models fail when familiar sub-tasks are recombined in new ways because they overfit to whole motions and to wrist-camera textures, and both failure modes can be removed by training on a mix of full demonstrations and automatically decomposed sub-task segments while masking wrist-view tokens during closed-gripper/place phases. On two compositional out-of-distribution suites, success rates rise by roughly 28 absolute points over the base policy, with in-distribution success nearly unchanged; a real-world arm shows a 47.5-point OOD gain. The reason to care is that VLA models otherwise cannot reuse skills like pick and place in unseen pairings, and simply adding more full trajectories is an inefficient fix.

Core claim

The paper's central claim is that trajectory overfitting and perceptual shortcuts are complementary, jointly responsible for out-of-distribution failures, and both are addressable at supervision time. An LLM decomposes each instruction into an ordered list of sub-task descriptions, a proprioceptive aligner splits the corresponding trajectory at gripper-state changes or low-displacement moments, and the policy is trained jointly on full demonstrations and these sub-task segments, while wrist-view tokens are masked whenever the gripper is closed. The claimed result is that in-distribution success stays near its original level while compositional OOD success jumps by roughly 28 absolute points

What carries the argument

The central mechanism is the proprioceptive-driven aligner: a boundary detector that flags a sub-task boundary whenever the gripper state changes (open to closed or closed to open) or the chunk-wise average end-effector displacement falls below a threshold, producing trajectory segments meant to match the LLM's ordered list of sub-task descriptions. The second mechanism is state-conditioned asymmetric masking, which suppresses wrist-view tokens in the attention computation only during closed-gripper phases. Together, with a mixed training schedule that interleaves full and decomposed data, these carry the argument without modifying the underlying policy architecture.

Load-bearing premise

The load-bearing premise is that the boundary detector (gripper-state change or near-zero motion) produces exactly as many segments, in the same order, as the LLM's sub-task list; the paper does not specify how these two counts are reconciled, so if they disagree the decomposed supervision is mislabeled and the reported OOD gains are not explained.

What would settle it

Hand-label the true pick/place boundaries on a sample of training demonstrations, then count how often the proprioceptive boundary detector produces the same number of segments as the LLM's sub-task descriptions and whether the boundaries align in order. If the counts disagree often, or if substituting random fixed-length chunks for the LLM-plus-aligner gives the same ~28-point OOD gain, the compositional mechanism would cease to be the operative cause.

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

If this is right

  • Any VLA backbone can adopt the recipe without new modules, so the training-time gains transfer beyond the base policy; a second generalist backbone shows a similar ~20-point OOD improvement.
  • Decomposed-only training is not enough: without mixing with full demonstrations, in-distribution success collapses from roughly 97% to about 61%, so the prescription includes mixing both granularities.
  • Masking alone, without decomposition, already lifts OOD success (from 35.5 to 47.3 and from 46.6 to 67.0 on the two suites), meaning wrist-shortcut removal is independently effective.
  • The effects compound: combining decomposition with masking produces the largest gains (64.2 and 73.3 on the two OOD suites), and attention in the place phase shifts away from the old target toward the instructed target.
  • Real-world OOD trials reproduce the pattern, with OOD success rising from 35% to 82.5% while in-distribution stays high, suggesting the mechanism transfers beyond simulation.

Where Pith is reading between the lines

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

  • Inference: the paper leaves unspecified how the number of LLM-determined sub-tasks is matched to the physical boundary count from the proprioceptive aligner, and what happens when they disagree (Section 3.2.1, Eq. 2). If that matching is loose, the dense sub-task supervision may contain mislabeled segments, so an ablation against random or fixed-length segment boundaries would isolate how much of
  • Inference: if the mechanism is truly compositional, OOD gains should grow with the diversity of sub-skills in the training data even when full-demonstration coverage is held fixed; this is a testable scaling prediction.
  • Inference: the wrist-masking result suggests a cheap data-level robustness fix for VLA policies: any task with recognizable grasp/place phases could apply state-conditioned masking at training time without full decomposition, since masking alone already improves OOD.

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. AC-VLA is a training-time, architecture-agnostic add-on for VLA policies. It has two components: (i) a compositional learning module that uses an LLM to decompose each task instruction into sub-task descriptions and a proprioceptive aligner (gripper-state transitions, low end-effector displacement) to segment the corresponding trajectory, then trains the policy on a mixture of full demonstrations and decomposed sub-task data; and (ii) a state-conditioned asymmetric masking strategy that masks wrist-view tokens during closed-gripper (place) phases. The paper instantiates the method on π0.5 and GR00T-N1 and reports ~28 pt absolute improvement on LIBERO-OOD Spatial/Goal suites (π0.5: 35.5→64.2, 46.6→73.3), while keeping in-distribution success near 96-98%, plus a real-world OOD gain (35.0→82.5). The authors identify two failure modes—trajectory overfitting and wrist-view perceptual shortcuts—and their ablations (Table 3) suggest both components contribute additively.

Significance. If the reported results are robust, AC-VLA would be a practically valuable plug-and-play method: it requires no architectural changes, addresses a clear OOD failure mode, and the Table 3 ablation gives a coherent causal story that raw training is OOD-fragile, mixed training adds ~16/21 points, masking adds ~12/20 points, and the combination gives +28.7/+26.7. The attention visualizations and real-world experiments are appropriate, and the in-distribution preservation is an important check. However, the manuscript currently lacks a precise specification of the central LLM-to-proprioception alignment, gives no error bars or code, and does not quantify decomposition quality; these gaps make the main claim not yet verifiable. The strengths are the large, internally consistent effect sizes and the honest limitation section, which itself flags the aligner-fidelity dependence that the experimental section does not measure.

major comments (3)
  1. [Section 3.2.1, Eq. (2)] The paper never specifies how the LLM's N sub-task descriptions {ℓ_n} are matched to the physical segments produced by seg(t). Eq. (1) writes s_i = (ℓ_n, ξ_n) and the text jumps to 'aligned pairs (ℓ_k, ξ_k)' forming D_sub, but no algorithm is given for the case where the number of detected boundaries differs from N, where gripper-state transitions occur multiple times within one semantic sub-task, or where the ordering of detected segments disagrees with the LLM's plan. Since D_sub is the claimed source of compositional generalization (Table 3) and is also used to locate 'place' phases for asymmetric masking (§3.3), a mismatch would mislabel supervision and make the OOD gain uninterpretable. Please specify the matching rule (or a robust assignment procedure) and report a quantitative validation of the segmentation/alignment (e.g., agreement with human annotations or with the LLM sub-task
  2. [Sections 3.2.1-3.2.2] Two control parameters that determine the method are never quantified. The displacement threshold ε is said to be 'determined via clustering on the training trajectories,' but the clustering algorithm and the resulting value(s) are not reported; the 'fixed ratio' for sampling mini-batches from D_full and D_sub is also not given. Moreover, the segmentation extraction rule is ambiguous: 'we segment the full trajectory in a chunk-wise manner; thus, a pick or place sub-task can be obtained whenever at least one segmentation point exists in a chunk'—if a boundary falls anywhere in a chunk, is the entire chunk the segment, and how are consistent start/end timestamps constructed? Without these details, D_sub cannot be reconstructed.
  3. [Sections 4.3-4.4, Tables 3-4] No measure of variance is reported even though §4.1 says each simulation task is evaluated over 3 seeds x 50 episodes. Table 3's in-distribution ablations (96.6 vs 96.5 vs 96.7) are likely within noise, and the real-world OOD result 35.0→82.5 is reported without stating the number of trials. Also, Section 6 acknowledges that decomposition accuracy depends on aligner fidelity, but no experiment measures that fidelity. Please add standard deviations, trial counts for the real-world evaluation, and a segmentation-quality statistic.
minor comments (5)
  1. [References] Reference [49] (OpenVLA-OFT) lists the same arXiv identifier as Ref [5]; please correct the entry.
  2. [Section 4.2] The text mentions 'GR00T-N1.7' as a baseline, but Table 2 and Reference [2] use 'GR00T-N1'; make the naming consistent.
  3. [Section 4.1] The LLM decomposer is named as Qwen-3.5-Flash, but no prompt template, decomposition examples, or hyperparameters are provided. This is needed for reproducibility.
  4. [Figure 4] The claim of a 'larger than 100% increase in per-token attention' is given without numeric values or error bars; please report the actual scores and the number of episodes averaged.
  5. [Table 2] The 'AVG' column is not defined in the caption; state explicitly that it is the mean success rate across the six reported suites.

Circularity Check

0 steps flagged

No significant circularity: the reported OOD gains are benchmarked externally, and no prediction reduces by construction to a fitted parameter or self-citation.

full rationale

AC-VLA is an empirical training-strategy paper rather than a derivation from first principles. The compositional learning module decomposes instructions with an LLM and aligns proprioceptive segments via Eq. (2); the epsilon threshold is obtained by clustering on training trajectories, but it enters only the boundary detector and is never fitted to the success-rate numbers that are later reported. The central claims are measured against external benchmarks (LIBERO, LIBERO-OOD, and real-world OOD tasks), so the OOD improvements are not defined into existence by the method's own equations. The one structurally suspicious item—the claim that LLM sub-task descriptions (ℓ_k) are paired one-to-one with proprioceptive segments (ξ_k)—is under-specified (the paper gives no matching rule when boundary counts disagree), and the Limitations section concedes that decomposition accuracy depends on aligner fidelity. However, under-specification and potential mislabeling are correctness risks, not circularity: even if the pairing failed, that would make the attribution unproven, not make the result equivalent by construction to an input. Reference [11] includes overlapping authors but is not load-bearing for the OOD claim. No renamed known result, no imported uniqueness theorem, and no ansatz smuggled via self-citation are present. Score 0.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central claim depends on the decomposition producing semantically correct paired data (epsilon and LLM-to-segment matching), on masking not removing needed information, and on the benchmark meaning what it claims. No new physical entities are introduced; two free parameters are hand-set or fitted choices.

free parameters (2)
  • epsilon displacement threshold for sub-task segmentation
    Eq. (2) defines a sub-task boundary when chunk-wise displacement D(A_t) < epsilon. Section 3.2.1 states epsilon is 'determined via clustering on the training trajectories.' It directly controls how demonstrations are split, hence the quality of D_sub.
  • ratio of full-demonstration to sub-task mini-batches in mixed training
    Section 3.2.2 says mini-batches are sampled from D_full and D_sub 'with a fixed ratio,' but the ratio is never stated. The ablation shows mixed training is the mechanism preserving in-distribution performance while gaining OOD, so this hand-set hyperparameter is load-bearing.
axioms (5)
  • domain assumption Gripper-state transitions and low chunk-wise displacement are semantic sub-task boundaries.
    Eq. (2) equates seg(t)=True with either g_t != g_{t-1} or D(A_t) < epsilon. The entire D_sub dataset depends on this kinematic proxy matching semantic phases such as 'pick' and 'place'.
  • ad hoc to paper LLM sub-goals and proprioceptive segments can be paired one-to-one and in order.
    N is 'dynamically determined by the LLM' while physical boundaries come from Eq. (2). The paper never states the matching rule or a failure-handling rule, making this an implicit ad hoc assumption.
  • domain assumption The chosen LLM (Qwen-3.5-Flash) decomposes instructions correctly enough for training.
    Acknowledged as a limitation in Section 6; the decomposition quality gates all sub-task supervision and trajectory alignment.
  • domain assumption Wrist-view information is not needed during closed-gripper/place phases for the downstream policy.
    Asymmetric masking removes wrist tokens during those phases. The results suggest the assumption mostly holds, but real-world ID drops from 93.7 to 88.7, so it is not guaranteed in general.
  • domain assumption LIBERO-OOD success rate measures compositional OOD generalization.
    The evaluation relies on the LIBERO-OOD suite from [47]. If those tasks are simple recombinations, the reported gains may not transfer to harder novel-configuration regimes.

pith-pipeline@v1.3.0-alltime-deepseek · 10063 in / 15959 out tokens · 127505 ms · 2026-08-01T22:31:10.224755+00:00 · methodology

0 comments
read the original abstract

Vision-Language-Action (VLA) models excel at end-to-end robotic manipulation but struggle with out-of-distribution (OOD) generalization when familiar sub-tasks are recombined in unseen configurations. We identify two mutually reinforcing failure modes: \emph{trajectory overfitting}, where models overfit to holistic trajectory patterns rather than compositional sub-skill semantics; and \emph{perceptual shortcut}, where action tokens over-rely on wrist-view textures at the expense of global spatial grounding. To address both, we introduce \textbf{AC-VLA}, a plug-and-play Action Compositional learning framework comprising two architecture-agnostic components: \textbf{(i)} a compositional learning module that uses an LLM-driven instruction decomposer and a proprioceptive trajectory aligner to generate dense sub-task supervision, followed by mixed training on complete demonstrations and decomposed data to endow the model with compositional generalization; and \textbf{(ii)} a state-conditioned asymmetric masking strategy that suppresses wrist-view inputs during closed-gripper phases, enforcing global semantic grounding. All components are architectural modification-free and directly integrable into any VLA backbone. Instantiated on $\pi_{0.5}$ and evaluated on LIBERO and LIBERO-OOD benchmarks, AC-VLA achieves a ~28% absolute improvement on compositional OOD tasks while maintaining near-perfect in-distribution performance.

Figures

Figures reproduced from arXiv: 2607.15714 by Jie Lu, Kai Peng, Xiaobo Wang, Xiaojiang Peng, Zheng Lian, Zitong Yu.

Figure 1
Figure 1. Figure 1: Out-of-distribution compositional actions confuse [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the AC-VLA framework. (a) Compositional learning module: full demon [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evidence of perception shortcuts in model atten [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of attention maps between the vanilla [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Successful real-world OOD examples. 4.4 Evaluation on Real-World Scene We further evaluate AC-VLA on a real-world robotic platform. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_6.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

49 extracted references · 14 linked inside Pith

  1. [1]

    W. Li, R. Zhang, R. Shao, J. He, and L. Nie. Cogvla: Cognition-aligned vision-language-action model via instruction-driven routing & sparsification. InAdvances in Neural Information Pro- cessing Systems (NeurIPS), 2025

    2025

  2. [2]

    Bjorck, F

    J. Bjorck, F. Casta ˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Black, N

    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.arXiv preprint arXiv:2410.24164, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Zitkovich, R

    A. Zitkovich, R. Jackson, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. InConference on Robot Learning (CoRL), 2023

    2023

  5. [6]

    J. Wen, Y . Zhu, J. Li, M. Zhu, Z. Tang, K. Wu, Z. Xu, N. Liu, R. Cheng, C. Shen, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. IEEE Robotics and Automation Letters, 2025

    2025

  6. [7]

    Shukor, D

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafioti, et al. Smolvla: A vision-language-action model for afford- able and efficient robotics.arXiv preprint arXiv:2506.01844, 2025

    Pith/arXiv arXiv 2025

  7. [8]

    Brohan, N

    A. Brohan, N. Brown, Y . Chebotar, et al. Rt-1: Robotics transformer for real-world control at scale. InRobotics: Science and Systems (RSS), 2023

    2023

  8. [9]

    Ghosh, H

    D. Ghosh, H. Walke, K. Pertsch, et al. Octo: An open-source generalist robot policy. In Robotics: Science and Systems (RSS), 2024

    2024

  9. [10]

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu. Rdt-1b: a diffu- sion foundation model for bimanual manipulation. InInternational Conference on Learning Representations (ICLR), 2025

    2025

  10. [11]

    Liang, X

    Y . Liang, X. Wang, K. Wang, S. Wang, X. Peng, H. Chen, D. K. H. Chua, and P. Vadakkepat. Adaptive action chunking at inference-time for vision-language-action models.arXiv preprint arXiv:2604.04161, 2026

    Pith/arXiv arXiv 2026

  11. [12]

    Y . Fang, Y . Feng, D. Jing, J. Liu, Y . Yang, Z. Wei, D. Szafir, and M. Ding. When vision overrides language: Evaluating and mitigating counterfactual failures in vlas.arXiv preprint arXiv:2602.17659, 2026

    Pith/arXiv arXiv 2026

  12. [13]

    F. Lin, Y . Hu, et al. Data scaling laws in imitation learning for robotic manipulation. InThe Thirteenth International Conference on Learning Representations (ICLR), 2024

    2024

  13. [14]

    Zhao et al

    Z. Zhao et al. Robobrain 2.0: A spatial-temporal reasoning world model for robotic manipula- tion.arXiv preprint arXiv:2506.18829, 2025. 10

    Pith/arXiv arXiv 2025

  14. [15]

    M. Ahn, A. Brohan, N. Brown, et al. Do as i can, not as i say: Grounding language in robotic affordances. InConference on Robot Learning (CoRL), 2022

    2022

  15. [16]

    Huang, F

    W. Huang, F. Xia, T. Xiao, et al. Inner monologue: Embodied reasoning through planning with language models. InConference on Robot Learning (CoRL), 2022

    2022

  16. [17]

    Driess, F

    D. Driess, F. Xia, M. S. M. Sajjadi, et al. Palm-e: An embodied multimodal language model. InInternational Conference on Machine Learning (ICML), 2023

    2023

  17. [18]

    Liu et al

    J. Liu et al. Long-vla: Scaling long-context vision-language-action models with phase-aware decomposition and masking.arXiv preprint arXiv:2502.16925, 2025

    Pith/arXiv arXiv 2025

  18. [19]

    Chen et al

    Z. Chen et al. Learning semantic atomic skills for multi-task robotic manipulation.arXiv preprint arXiv:2512.18368, 2024

    Pith/arXiv arXiv 2024

  19. [20]

    X. Li, M. Liu, H. Zhang, et al. Vision-language foundation models as effective robot imitators. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  20. [21]

    D. Qu, H. Ge, Z. Shi, et al. Spatialvla: Exploring spatial representations for visual-language- action model.arXiv preprint arXiv:2501.12905, 2025

    Pith/arXiv arXiv 2025

  21. [22]

    Doshi, H

    R. Doshi, H. Walke, O. Mees, et al. Scaling cross-embodied learning: One policy for ma- nipulation, navigation, locomotion, and aviation. InConference on Robot Learning (CoRL), 2024

    2024

  22. [23]

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023

    Pith/arXiv arXiv 2023

  23. [24]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, et al. Learning transferable visual models from natural language supervision. InInternational Conference on Machine Learning (ICML), 2021

    2021

  24. [25]

    H. Liu, C. Li, Q. Wu, and Y . J. Lee. Visual instruction tuning. InAdvances in Neural Informa- tion Processing Systems (NeurIPS), 2024

    2024

  25. [26]

    Alayrac, J

    J.-B. Alayrac, J. Donahue, P. Luc, et al. Flamingo: A visual language model for few-shot learning. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  26. [27]

    Padalkar, A

    Open X-Embodiment Collaboration, A. Padalkar, A. Pooley, et al. Open X-Embodiment: Robotic learning datasets and RT-X models. InIEEE International Conference on Robotics and Automation (ICRA), 2024

    2024

  27. [28]

    Khazatsky, K

    A. Khazatsky, K. Tajwar, et al. DROID: A large-scale in-the-wild robot manipulation dataset. InRobotics: Science and Systems (RSS), 2024

    2024

  28. [29]

    B. Liu, Y . Jiang, X. Zhang, S. Bahl, P. Stone, and Y . Zhu. LIBERO: Benchmarking knowledge transfer in lifelong robot learning. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  29. [30]

    Liang, W

    J. Liang, W. Huang, F. Xia, et al. Code as policies: Language model programs for embodied control. InIEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  30. [31]

    Black, M

    K. Black, M. Nakamoto, et al. Zero-shot robotic manipulation with pre-trained image-editing diffusion models. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  31. [32]

    H. Chen, J. Guo, et al. Goal-vla: Image-generative vlms as object-centric world models em- powering zero-shot robot manipulation, 2025

    2025

  32. [33]

    Kang and Y .-L

    X. Kang and Y .-L. Kuo. Incorporating task progress knowledge for subgoal generation in robotic manipulation through image edits. In2025 IEEE/CVF Winter Conference on Applica- tions of Computer Vision (WACV). IEEE, 2025. 11

    2025

  33. [34]

    Singh, V

    I. Singh, V . Blukis, A. Mousavian, et al. Progprompt: Generating situated robot task plans using large language models. InIEEE International Conference on Robotics and Automation (ICRA), 2023

    2023

  34. [35]

    G. Wang, Y . Xie, Y . Jiang, et al. V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291, 2023

    Pith/arXiv arXiv 2023

  35. [36]

    D. Garg, S. Vaidyanath, K. Kim, et al. LISA: Learning interpretable skill abstractions from language. InAdvances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  36. [37]

    L. Chen, K. Lu, A. Rajeswaran, et al. Decision transformer: Reinforcement learning via sequence modeling. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  37. [38]

    Q. Zhao, Y . Lu, M. J. Kim, Z. Fu, Z. Zhang, Y . Wu, Z. Li, Q. Ma, S. Han, C. Finn, et al. Cot- vla: Visual chain-of-thought reasoning for vision-language-action models. InProceedings of the Computer Vision and Pattern Recognition Conference, 2025

    2025

  38. [39]

    C. Wang, L. Fan, J. Sun, et al. Mimicplay: Long-horizon imitation learning by watching human play. InConference on Robot Learning (CoRL), 2023

    2023

  39. [40]

    Bharadhwaj, J

    H. Bharadhwaj, J. Vakil, M. Sharma, et al. Roboagent: Generalization and efficiency in robot manipulation via semantic augmentation. InIEEE International Conference on Robotics and Automation (ICRA), 2024

    2024

  40. [41]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware. InRobotics: Science and Systems (RSS), 2023

    2023

  41. [42]

    C. Chi, Z. Xu, C. Pan, et al. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. InRobotics: Science and Systems (RSS), 2024

    2024

  42. [43]

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. InRobotics: Science and Systems (RSS), 2023

    2023

  43. [44]

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

    2024

  44. [45]

    J. Luo, C. Xu, J. Wu, and S. Levine. Precise and dexterous robotic manipulation via human- in-the-loop reinforcement learning.Science Robotics, 10(105):eads5033, 2025

    2025

  45. [46]

    Hejna et al.π 0.5: a vision-language-action model with open-world generalization.arXiv preprint arXiv:2504.16054, 2025

    J. Hejna et al.π 0.5: a vision-language-action model with open-world generalization.arXiv preprint arXiv:2504.16054, 2025

    Pith/arXiv arXiv 2025

  46. [47]

    Q. Li. Vlas are confined yet capable of generalizing to novel instructions. InIJCAI, 2026

    2026

  47. [48]

    Bu et al

    Q. Bu et al. Univla: Learning to act anywhere with task-centric latent actions.arXiv preprint arXiv:2505.06111, 2025

    Pith/arXiv arXiv 2025

  48. [49]

    M. J. Kim, K. Pertsch, S. Karamcheti, et al. Openvla-oft: Fine-tuning vision-language-action models via optimal transport with action chunking.arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  49. [50]

    F. Li, W. Song, H. Zhao, J. Wang, P. Ding, D. Wang, L. Zeng, and H. Li. Spatial forcing: Implicit spatial representation alignment for vision-language-action model.arXiv preprint arXiv:2510.12276, 2025. 12

    arXiv 2025