Pith. sign in

REVIEW 3 major objections 5 minor 10 cited by

Long-VLA: Unleashing Long-Horizon Capability of Vision Language Action Model for Robot Manipulation

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Long-VLA claims that an end-to-end vision-language-action policy can solve long-horizon manipulation by masking its input according to whether the subtask is in a moving or an interaction phase.

desk verdict Plausible idea with a clean base-policy control, but inconsistent numbers and a data-confounded SOTA comparison make the current version unready. read the letter →

arxiv 2508.19958 v2 pith:66TGO3YI submitted 2025-08-27 cs.RO

Yiguo Fan , Pengxiang Ding , Shuanghao Bai , Xinyang Tong , Yuyang Zhu , Hongchao Lu , Fengqi Dai , Wei Zhao
show 5 more authors
Yang Liu Siteng Huang Zhaoxin Fan Badong Chen Donglin Wang
This is my paper
classification cs.RO
keywords vision-language-actionmodelslong-horizonmanipulationskillchainingphase-awaremaskingdiffusionpolicyL-CALVINbenchmarkrobot
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

The paper's central claim is that vision-language-action models fail at long-horizon manipulation not because individual skills are weak but because of skill chaining: errors and visual distribution shifts accumulate across subtask boundaries. Long-VLA addresses this by splitting each subtask into a moving phase and an interaction phase and applying a phase-aware attention mask that lets a single end-to-end diffusion policy attend to the third-person camera during navigation and the gripper camera during contact. The design is meant to preserve the scalability and data efficiency of end-to-end VLA training while avoiding the two-model decomposition and online reward loops used by earlier long-horizon methods. On the new 10-task L-CALVIN benchmark and on real-world sorting and cleaning tasks, the paper reports average completed-task lengths and success rates well above prior state-of-the-art baselines, with the gap widening as sequences get longer. If the claim holds, long-horizon manipulation can be improved by a cheap input-level adaptation rather than a new planner or hierarchical controller.

What carries the argument

Phase-aware input masking. Each token is assigned a binary mask m_i; the attention mask matrix is M_ij = m_i * m_j, so attention is computed only between tokens that are both active in the current phase. A phase identifier s_p in the action token tells the model which phase is active (-1 moving, +1 interaction), switching the policy between a third-person 'where to go' mode and a gripper-camera 'how to interact' mode. This mechanism is what lets one unified transformer replace two separately trained policies while still addressing the skill-chaining problem.

What would settle it

Corrupt the phase signal during L-CALVIN rollouts—flip s_p randomly every few steps, or force it to -1 throughout the interaction phase. If average completed-task length stays near Long-VLA's 4.75, phase timing is not what carries the result; if it falls toward the base policy's 4.11, the phase switch is load-bearing and needs an explicit detector.

Watch

Extended reading notes

Core claim

Long-VLA claims that phase-aware input masking is sufficient to substantially improve long-horizon manipulation within an end-to-end VLA. During training, each trajectory is cut into a moving phase and an interaction phase; the action token carries a phase identifier s_p = -1 for movement and +1 for interaction. The mask assigns each input token a binary flag, and attention is restricted to pairs of active tokens via M_ij = m_i * m_j, so third-person static-camera tokens drive the moving phase while gripper-camera tokens dominate interaction. This replaces the previous two-stage decomposition (separate moving and interaction policies) with one jointly trained model. On L-CALVIN, Long-VLA rea

Load-bearing premise

The load-bearing premise is that the robot knows at execution time when the moving phase ends and the interaction phase begins; the paper initializes the phase signal to -1 at inference but does not state how the +1 switch is detected.

Editorial extensions

If this is right

  • Long-horizon tasks can be chained in a single end-to-end diffusion policy: no separate planner, no online reward correction, and no modular policy library are needed for the reported gains.
  • The module is architecture-agnostic: applying the same masking and phase conditioning to a different backbone (HULC) raises its average completed length from 2.65 to 3.30, just as it raises MDT from 4.11 to 4.81.
  • Performance gains grow with horizon and visual distraction: in real-world sorting, the base policy completes zero of eight tasks while Long-VLA completes all eight with 45% success under unseen localization changes, and the cleaning task shows even larger relative improvements.
  • The L-CALVIN benchmark's 10-task protocol and phase-split sequences separate models that look similar on CALVIN's 5-task protocol, giving future VLA comparisons a stricter test.

Reading between the lines

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

  • The paper leaves the inference-time phase switch unspecified—it only states s_p starts at -1—so the cleanest extension is a learned phase-boundary predictor; if phase timing is accurate, the same masking should transfer to tasks where manually cutting trajectories is infeasible.
  • The same input-level masking idea could be applied to other modality mismatches in robot learning—e.g., reach-versus-grasp in bimanual or mobile manipulation—where a mask is learned rather than fixed by camera view.
  • The paper's own limitation section concedes that phase decomposition of training data is manual and that failure cases in longer sequences are not handled; a testable next step is to replace manual cutting points with an automatic detector and measure how L-CALVIN success degrades as cutting-point error grows.
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

3 major / 5 minor

Summary. Long-VLA proposes a unified end-to-end vision-language-action (VLA) model for long-horizon robot manipulation. The method decomposes each subtask into a moving phase and an interaction phase, and applies hand-designed, phase-aware attention masking: third-person camera tokens are used in the moving phase and gripper-camera tokens in the interaction phase. A one-dimensional phase identifier is appended to the action representation, and a diffusion action head is trained with a combined score-matching and contrastive goal loss. The paper also introduces L-CALVIN, a 10-task extension of the CALVIN benchmark, and reports substantial gains over the MDT base policy and over prior SOTA methods (GR-1, RoboVLMs, VLAS, etc.) in simulation and in two real-world tasks (sorting and cleaning).

Significance. If substantiated, the central idea of phase-aware input masking is simple, architecture-agnostic, and potentially useful for mitigating skill-chaining errors in long-horizon manipulation. The L-CALVIN benchmark, if released with reproducible generation code, could also be a valuable community resource. The paper includes real-world experiments, a multi-backbone scalability study, and a failure-cause analysis, which are strengths. However, the current manuscript has a load-bearing gap in the specification of inference-time phase switching, internally inconsistent headline numbers, and SOTA comparisons that omit the paper's own base policy. These issues prevent the reader from verifying the central empirical claim.

major comments (3)
  1. [Section 3.2.1; Figure 2(b)] Inference-time phase switching is unspecified. The text states that s_p is set to -1 during the moving phase and +1 during the interaction phase, and that 'during inference, s_p is initialized to -1' (Section 3.2.1). No mechanism is described for changing s_p to +1 at the moving-to-interaction boundary. Figure 2(b) labels a 'Moving Phase Detection' block, but no algorithm is given. Since the attention mask depends on the phase, a policy that never switches would always use the moving-phase mask (static camera only, gripper camera masked), which would severely impair interaction-phase actions. The reported gains over the base policy on L-CALVIN and real-world tasks depend on this switch being reliable. Please specify the detector, or clarify whether s_p is a predicted action dimension and how the predicted value controls the mask, together with training/inference details. This is a load-b
  2. [Figure 4; Tables 2, 3, 4] The numerical results are internally inconsistent. In Figure 4 (D->D), the Base Policy per-step success rates sum to 0.86+0.64+0.53+0.47+0.37+0.31+0.28+0.21+0.13+0.11 = 3.91, but Table 3 and Table 4 report the same MDT base policy as Avg. Len 4.11. For Long-VLA (D->D), Figure 4 sums to 4.75, matching Table 2, but Table 3 and Table 4 report 4.81. Rounding cannot account for a 0.20 discrepancy in the base policy. The same experimental setting cannot have two different average lengths. Please reconcile these numbers and report per-step success counts or confidence intervals.
  3. [Table 2; Table 8] The SOTA comparison omits the paper's own base policy. By the paper's Figure 4 numbers, the MDT base policy on L-CALVIN D->D has average length 3.91, which is already higher than GR-1 (2.96) and RoboVLMs (2.88). Tables 2 and 8 do not include a Base Policy (MDT) row, so the reader cannot tell whether Long-VLA's advantage over these baselines is due to the proposed phase-aware mechanism or to the stronger MDT backbone. Please add the base policy row to all SOTA comparisons, or otherwise provide a controlled comparison.
minor comments (5)
  1. [Abstract] Typo: 'Long-WLA' should be 'Long-VLA'.
  2. [References and inline citations] Several references are corrupted: '79' and 'aq' and 'zo' appear to refer to π0 (e.g., Section 4.1 and Appendix D.1). Please fix these and any other mangled citations (e.g., '7' in Section 4.3).
  3. [Section 3.2.1] Unresolved cross-reference: 'Details of the masking strategy are provided in Appendix ??' — the appendix number is missing.
  4. [Table 3] Table 3 formatting is garbled: the row/column structure is hard to parse, and the entries like '3.60131' and '1.7(0.3 4)' are unclear. Please reformat and define the parenthetical values.
  5. [Section 3.2.2] The detection input d^t is used in the policy definition and Equation (2) but is not defined until the Detection Integration paragraph. Please define it at first use.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the central claims are empirical and the derivation chain is self-contained.

full rationale

The paper's central claims are empirical comparisons on the proposed L-CALVIN benchmark and on real-world tasks, not derivations from the method's definitions. The phase-aware masking design is a hand-chosen architectural choice whose value is established by ablations (Tables 3, 6, and the learnable-mask consistency check in Appendix D.5), rather than by an equation that reduces a prediction to its input. The learnable-mask experiment reports that the learned masks align with the design, but this is a post-hoc consistency observation and is not used as the evidence for the main performance claim. Citations to prior work such as Plan-Seq-Learn and RH20T-P are external, and the paper's many self-citations appear only as related work or background, not as load-bearing justification for the method's effectiveness. The inference-time phase-switch mechanism is under-specified (the text only says that s_p is initialized to -1), which is a reproducibility gap, but it is not a circularity: no construction makes the reported improvements equivalent to the training labels or to a fitted parameter. Therefore the derivation chain is not circular.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The central claim rests on phase labels built with manual cut points and a task detector, a hand-chosen mask design, and unstated runtime phase detection. The diffusion and attention machinery are standard math; the benefit of the phase mask is a domain assumption rather than a derivation.

free parameters (5)
  • Phase cutting point offset = 10-15 frames before object state change
    Hand-set window for splitting trajectories into moving and interaction phases; directly determines the phase labels used to train the phase token (Section 3.1, Appendix B.1).
  • Goal loss weight alpha = 0.1
    Weight balancing the diffusion loss and the InfoNCE goal loss in Eq. 3; no sensitivity analysis is reported.
  • Trajectory window size = 64 frames
    Window length used to extract L-CALVIN sequences (Appendix B.1).
  • Phase mask configuration = static camera masked in interaction; gripper camera masked in moving
    Hand-designed masking assignment (Section 3.2.1); ablations in Table 6 confirm it on this dataset but do not justify it more generally.
  • Phase identifier values = -1 (moving), +1 (interaction)
    Arbitrary scalar encoding appended to the action token (Section 3.2.1).
assumptions (6)
  • standard math Attention masking with M_ij = m_i * m_j produces valid attention weights (Eq. 1)
    Standard masked attention from the transformer literature, used without modification.
  • standard math Diffusion score-matching on actions (Eq. 2) with DDIM sampling generates demonstrated actions
    Standard diffusion policy machinery adopted from prior work.
  • domain assumption The CALVIN task detector reliably marks task completion and object state changes
    Used to build phase labels and L-CALVIN annotations (Section 3.1, Appendix B.1).
  • domain assumption The gripper camera is uninformative during movement and the static camera is distracting during interaction
    Central motivation for the phase mask; asserted in Section 3.2.1 but not measured.
  • domain assumption Grounding DINO fine-tuned with LoRA provides accurate bounding boxes for navigation and interaction
    Detection inputs are fused into static-camera features via FiLM and used at inference (Appendix C.1).
  • domain assumption Phase boundaries can be detected at inference time
    The phase token must switch during rollout, but the switching rule is never specified (Section 3.2.1).
invented entities (2)
  • Phase identifier token s_p
    purpose: Condition the policy on moving vs interaction phase and gate input masking
    Introduced in this paper; no external falsifiable handle beyond the reported task success rates.
  • L-CALVIN benchmark
    purpose: 10-step long-horizon evaluation benchmark derived from CALVIN with phase-decomposed annotations
    Constructed by the authors with manual movement-instruction templates and structured sequence generation; no release code or URL is provided, so other groups cannot yet use it independently.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Long-VLA: Unleashing Long-Horizon Capability of Vision Language Action Model for Robot Manipulation." pith.science (2026). https://pith.science/paper/66TGO3YI

@misc{pith2026250819958,
  author       = {Pith},
  title        = {Pith review of: Long-VLA: Unleashing Long-Horizon Capability of Vision Language Action Model for Robot Manipulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/66TGO3YI}},
  note         = {Machine review of arXiv:2508.19958}
}
read the original abstract

Vision-Language-Action (VLA) models have become a cornerstone in robotic policy learning, leveraging large-scale multimodal data for robust and scalable control. However, existing VLA frameworks primarily address short-horizon tasks, and their effectiveness on long-horizon, multi-step robotic manipulation remains limited due to challenges in skill chaining and subtask dependencies. In this work, we introduce Long-VLA, the first end-to-end VLA model specifically designed for long-horizon robotic tasks. Our approach features a novel phase-aware input masking strategy that adaptively segments each subtask into moving and interaction phases, enabling the model to focus on phase-relevant sensory cues and enhancing subtask compatibility. This unified strategy preserves the scalability and data efficiency of VLA training, and our architecture-agnostic module can be seamlessly integrated into existing VLA models. We further propose the L-CALVIN benchmark to systematically evaluate long-horizon manipulation. Extensive experiments on both simulated and real-world tasks demonstrate that Long-VLA significantly outperforms prior state-of-the-art methods, establishing a new baseline for long-horizon robotic control.

Discussion (0). Sign in to comment.

Forward citations

Cited by 10 Pith papers

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

  1. PAMAE: Phase-Aware-MoE Action Experts Towards Reliable Flow-Matching Vision-Language-Action Policies

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    PAMAE adds a phase-aware router and expert mixture to flow-matching VLA models, yielding up to 9.2% higher task success on multi-stage manipulation simulations via two-stage training.

  2. UniviewVLA: A Unified Multiview Vision-Language-Action Model with World Modeling

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    UniviewVLA generates multiview future views from two cameras via world modeling, plus token compression and view selection, to boost occlusion handling in robot manipulation while matching standard benchmark performance.

  3. $\mu$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...

  4. FiberTune: Preserving Action-Fiber Visual Residuals in Vision-Language-Action Fine-Tuning

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    FiberTune is a new fine-tuning objective that preserves action-fiber visual residuals in VLA policies, yielding performance gains on simulation and physical robot tasks.

  5. TriRelVLA: Triadic Relational Structure for Generalizable Embodied Manipulation

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    TriRelVLA introduces triadic object-hand-task relational representations and a task-grounded graph transformer with a relational bottleneck to improve generalization in robotic manipulation across scenes, objects, and tasks.

  6. World Pilot: Steering Vision-Language-Action Models with World-Action Priors

    cs.RO 2026-06 unverdicted novelty 5.0 of 10

    World Pilot augments VLA policies with world-action priors through latent and action steering pathways, reporting 84.7% success on LIBERO-Plus zero-shot OOD and top real-robot results across four tasks.

  7. Make Your VLA More Robust Without More Data By Interleaving Motion Planning

    cs.RO 2026-05 unverdicted novelty 5.0 of 10

    MPVI interleaves model-based motion planning with VLAs via VLM completion checking to achieve 113% higher task progress on BEHAVIOR-1K without extra data.

  8. QDepth-VLA: Quantized Depth Prediction as Auxiliary Supervision for Vision-Language-Action Models

    cs.CV 2025-10 conditional novelty 5.0 of 10

    Adding an auxiliary quantized-depth-token prediction task to a VLA policy improves manipulation success rates on LIBERO, Simpler, and real-robot pick-and-place tasks versus the open-pi-zero baseline.

  9. Unleashing More Actions via Action Compositional Training for VLA Models

    cs.RO 2026-07 unverdicted novelty 4.0 of 10

    ACT-VLA synthesizes novel demonstrations from existing VLA tasks via latent representations to reduce overfitting and improve generalization on manipulation tasks in simulation.

  10. Safe Embodied AI for Long-horizon Tasks: A Cross-layer Analysis of Robotic Manipulation

    cs.RO 2026-06 unverdicted novelty 3.0 of 10

    A structured literature survey of safety mechanisms in long-horizon robotic manipulation organized by intervention timing and strength of supporting evidence.

Reference graph

Works this paper leans on

25 extracted references · 18 canonical work pages · cited by 10 Pith papers

  1. [1]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. San- keti, Q. Vuong, et al. Openvla: An open-source vision-language-action model. In 8th Annual Conference on Robot Learning, 2024. (2] J. Duan, W. Yuan, W. Pumacay, Y. R. Wang, K. Ehsani, D. Fox, and R. Krishna. Manipulate-anything: Automating real-wo...

    work page 2024

  2. [3]

    Dalal, T

    M. Dalal, T. Chiruvolu, D. S. Chaplot, and R. Salakhutdinov. Plan-seq-learn: Language model guided rl for solving long horizon robotics tasks. In /nternational Conference on Learning Representations, 2024

    work page 2024

  3. [4]

    Brohan, N

    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. In Robotics: Science and Systems, 2023

    work page 2023

  4. [5]

    Zitkovich, T

    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. In Conference on Robot Learning, pages 2165-2183. PMLR, 2023

    work page 2023

  5. [6]

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, et al. Octo: An open-source generalist robot policy. In Robotics: Science and Systems, 2024

    work page 2024

  6. [7]

    Zhang, P

    H. Zhang, P. Ding, S. Lyu, Y. Peng, and D. Wang. Gevrm: Goal-expressive video generation model for robust visual manipulation. arXiv preprint arXiv:2502.09268, 2025

    arXiv 2025

  7. [8]

    M. J. Kim, C. Finn, and P. Liang. Fine-tuning vision-language-action models: Optimizing speed and success, 2025. URL https: //arxiv.org/abs/2502. 19645

    work page 2025

  8. [9]

    Y. Tian, S. Yang, J. Zeng, P. Wang, D. Lin, H. Dong, and J. Pang. Predictive inverse dynamics models are scalable learners for robotic manipulation. In Jnternational Conference on Learning Representations, 2025

    work page 2025

Show all 25 references
  1. [10]

    P. Ding, J. Ma, X. Tong, B. Zou, X. Luo, Y. Fan, T. Wang, H. Lu, P. Mo, J. Liu, et al. Humanoid-vla: Towards universal humanoid control with visual integration. arXiv preprint arXiv:2502.14795, 2025

    2025 arXiv

  2. [11]

    W. Song, J. Chen, P. Ding, Y. Huang, H. Zhao, D. Wang, and H. Li. Ceed-vla: Consistency vision- language-action model with early-exit decoding, 2025. URL https: //arxiv. org/abs/2506. 13725

    2025

  3. [12]

    W. Zhao, G. Li, Z. Gong, P. Ding, H. Zhao, and D. Wang. Unveiling the potential of vision-language- action models with open-ended multimodal instructions. 2025. URL https: //api.semanticscholar. org/Corpus ID : 278714573. [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] ...

    2025 arXiv

  4. [50]

    Konidaris, S

    G. Konidaris, S. Kuindersma, R. Grupen, and A. Barto. Robot learning from demonstration by construct- ing skill trees. The International Journal of Robotics Research, 31(3):360—375, 2012

    2012

  5. [51]

    M. Sun, P. Ding, W. Zhang, and D. Wang. Score-based diffusion policy compatible with reinforcement learning via optimal transport. ArXiv, abs/2502.12631, 2025. URL https://api.semanticscholar. org/Corpus ID :276421829

    2025 arXiv

  6. [52]

    Reuss, Omer Erding Ya&Smurlu, F

    M. Reuss, Omer Erding Ya&Smurlu, F. Wenzel, and R. Lioutikov. Multimodal diffusion transformer: Learning versatile behavior from multimodal goals. In Robotics: Science and Systems, 2024

    2024

  7. [53]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016

    2016

  8. [54]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In /nternational conference on machine learning, pages 8748-8763. PmLR, 2021

    2021

  9. [55]

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024

    2024

  10. [56]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al. Lora: Low-rank adaptation of large language models. International Conference on Learning Representations, 1(2):3, 2022

    2022

  11. [57]

    Perez, F

    E. Perez, F. Strub, H. De Vries, V. Dumoulin, and A. Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  12. [58]

    H. Wu, Y. Jing, C. Cheang, G. Chen, J. Xu, X. Li, M. Liu, H. Li, and T. Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. In The Twelfth International Conference on Learning Representations, 2024

    2024

  13. [59]

    Zhang, Y

    J. Zhang, Y. Guo, Y. Hu, X. Chen, X. Zhu, and J. Chen. Up-vla: A unified understanding and prediction model for embodied agent, 2025

    2025

  14. [60]

    X. Li, P. Li, M. Liu, D. Wang, J. Liu, B. Kang, X. Ma, T. Kong, H. Zhang, and H. Liu. Towards generalist robot policies: What matters in building vision-language-action models. arXiv preprint arXiv:2412.14058, 2024

    2024 arXiv

  15. [61]

    Z. Chen, Z. Shi, X. Lu, L. He, 8. Qian, Z. Yin, W. Ouyang, J. Shao, Y. Qiao, C. Lu, et al. Rh20t-p: A primitive-level robotic dataset towards composable generalization agents. arXiv preprint arXiv:2403.19622, 2024

    2024 arXiv

  16. [62]

    Put the red block or 1) cal into the drawer

    O. Mees, L. Hermann, and W. Burgard. What matters in language conditioned robotic imitation learning over unstructured data. JEEE Robotics and Automation Letters (RA-L), 7(4):11205—11212, 2022. A Preliminaries A.1l_ Definition of VLA Models Imitation learning with language ins...

    2022

  17. [367]

    Springer, 2024. X. Tong, P. Ding, D. Wang, W. Zhang, C. Cui, M. Sun, Y. Fan, H. Zhao, H. Zhang, Y. Dang, $. Huang, and S. Lyu. Quart-online: Latency-free large multimodal language model for quadruped robot learning. In EEE International Conference on Robotics and Automation, 2...

    2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.