Pith. sign in

REVIEW 4 major objections 4 minor 45 references

The paper argues that robotic manipulation evaluation should include fine-grained execution quality and source authenticity, and provides a benchmark and evaluator that achieve expert-level quality ranking.

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

2026-08-03 07:54 UTC pith:PDBKIAAY

load-bearing objection A genuinely useful benchmark paper whose central numbers rest on an under-reported expert ground truth; worth refereeing seriously, but the authors need to show inter-rater reliability and fix the checkpoint-selection ambiguity. the 4 major comments →

arxiv 2601.18723 v2 pith:PDBKIAAY submitted 2026-01-26 cs.RO

Eval-Actions: Fine-Grained Execution Quality Evaluation for Robotic Manipulation

Mengyuan Liu , Juyi Sheng , Peiming Li , Ziyi Wang , Tianming Xu , Tiantian Xu , Hong Liu This is my paper
classification cs.RO
keywords robotic manipulation evaluationaction quality assessmentvision-language-action modelsbenchmark datasetimitation learningexecution qualitysource authenticitychain-of-thought reasoning
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 tries to establish that binary success rates are insufficient for judging robotic manipulation policies, and that fine-grained execution quality — smoothness, safety, efficiency — can be formalized and learned. It introduces Eval-Actions, a 13k-episode real-robot benchmark with failure cases, mixed human/policy trajectories, and three kinds of expert supervision: numeric grades, rank-guided preferences, and chain-of-thought reasoning. On top of it, AutoEval predicts quality scores, success, and trajectory source from video plus kinematic summaries, reaching 0.84 rank correlation with expert rankings and 99.6% source accuracy. If right, evaluation of robot policies becomes a diagnostic instrument instead of a pass/fail gate, revealing shaky or unsafe execution and exposing teleoperation masquerading as autonomous behavior.

Core claim

Eval-Actions and AutoEval together claim that the trustworthiness of a robot policy can be assessed along three measurable axes — task success, fine-grained execution quality (smoothness, safety, efficiency), and source authenticity (policy vs. human teleoperation) — and that a vision-language evaluator can predict all three from RGB frames plus lightweight kinematic statistics. On the EAS test split, AutoEval-S reaches 0.81 Spearman correlation with expert grades, 0.84 with rank-guided labels, 91.0% success detection, and 99.6% teleoperation-vs-policy source accuracy; AutoEval-P reaches 0.70 with chain-of-thought explanations. The paper concludes that such diagnostic signals are standardize

What carries the argument

The carrying mechanism is a three-signal annotation stack — Expert Grading, Rank-Guided preferences, Chain-of-Thought reasoning — treated as ground truth, paired with the AutoEval architecture. AutoEval-S's Spatio-Temporal Aggregation stitches intermediate frames around each keyframe into composite visual tokens so high-frequency motion defects survive the token budget; an auxiliary Kinematic Calibration Signal (maximum per-joint variance of angular velocity and acceleration, plus mean absolute velocity) is serialized into a text prompt to stabilize smoothness scoring. AutoEval-P wraps the same inputs in Group Relative Policy Optimization to enforce logical consistency between reasoning toke

Load-bearing premise

Human experts' grades and rankings are treated as reliable ground truth for execution quality, yet inter-rater agreement is not reported in the main text; if expert judgment is noisy or inconsistent, every SRCC measures agreement with an unstable reference.

What would settle it

Two checks. (1) Re-score a random sample of EAS episodes by the 10 experts and compute inter-rater reliability (e.g., intraclass correlation on EG scores); if agreement is markedly below the model-to-expert correlation, the reported 0.81–0.84 SRCCs are inflated relative to a stable standard. (2) Feed the source classifier policy videos deliberately smoothed to mimic teleoperation statistics; if 99.6% accuracy falls sharply, the authenticity signal is a kinematic shortcut rather than a trustworthy provenance check.

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

If this is right

  • Policies that complete tasks with jitter, collisions, or inefficiency can now be separated from clean performers even when both receive success score 1.
  • A benchmark with 2.8k failure episodes and mixed teleoperation/policy trajectories provides training signal for failure detection and provenance verification.
  • AutoEval-S's 0.84 SRCC under rank-guided labels means automated evaluators can rank-order policy executions consistently with experts, enabling offline policy comparison.
  • AutoEval-P's chain-of-thought outputs provide human-readable justifications for scores, reducing hallucination-driven over-scoring.
  • The framework transfers to an unseen embodiment with SRCC 0.71–0.75, suggesting the quality assessment is not tied to one robot's visual appearance.

Where Pith is reading between the lines

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

  • The same scores could be inverted into a training reward, so policies could be optimized directly for predicted smoothness and safety rather than only task completion — a testable extension the paper does not run.
  • If expert annotations are as reliable as claimed, the EG/RG protocols could be adapted as a standardized robotics action-quality benchmark across labs, but that depends on inter-rater agreement measures the main text defers to the supplementary material.
  • The near-perfect source prediction may be vulnerable to policies trained to imitate teleoperation smoothness; adversarial provenance tests would clarify whether authenticity is genuinely robust.
  • CoT scoring at 0.70 SRCC suggests reasoning quality comes at a numeric cost; a future model that separates the reasoning tokens from the score head might get both interpretability and precision.

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

4 major / 4 minor

Summary. The paper introduces Eval-Actions, a benchmark for fine-grained execution-quality evaluation of robotic manipulation policies, and AutoEval, a VLM-based evaluator. The benchmark provides 13k+ real-robot episodes across 150+ tasks with RGB-D video, kinematic trajectories, failure cases, hybrid human/policy sources, and three annotation types: Expert Grading (EG), Rank-Guided preferences (RG), and Chain-of-Thought (CoT) reasoning. AutoEval-S uses spatio-temporal frame aggregation and a kinematic prompt to predict quality scores, success, and source; AutoEval-P applies GRPO for CoT generation. On the EAS test split, the paper reports SRCC 0.81 under EG, 0.84 under RG, success detection accuracy around 91%, near-perfect source prediction (99.6%), and CoT SRCC 0.70, claiming state-of-the-art diagnostic evaluation.

Significance. If the reported numbers and annotations are reliable, the benchmark addresses a real gap: binary success rates ignore execution smoothness, safety, and provenance. The inclusion of failure scenarios, policy-vs-teleoperation trajectories, dense EG/RG/CoT labels, and the AutoEval reference evaluator would be a useful community resource, especially if the code and data are released as claimed. The paper also provides several ablations and a cross-embodiment test, which are strengths. However, the central premise depends on the validity of expert annotations as ground truth, and that validity is not established in the main text. The small SRCC gaps over strong baselines also need statistical support. These issues make the current claims plausible but not yet publication-ready.

major comments (4)
  1. [Section III-B, Abstract, Section V]
  2. [Section V-B]
  3. [Section III-B and Section IV-A, Eq. (1)-(3)]
  4. [Table II and Section V-C]
minor comments (4)
  1. [Section V-D and Table III]
  2. [Section V-C and Figure 7]
  3. [Section III-C]
  4. [Various]

Circularity Check

0 steps flagged

No significant circularity — RG is transparent label generation; EG/CoT provide independent human supervision; the single MP1 self-citation is not load-bearing.

full rationale

The paper's central predictions (EG SRCC 0.81, RG SRCC 0.84, CoT SRCC 0.70, source 99.6%) are obtained by supervised fine-tuning on held-out expert annotations, and no reported prediction reduces to a fitted parameter of the same equation. The RG labels are constructed by GA weight optimization (Eqs. 1-3) to align kinematic metrics with expert rankings; this is label generation, not the evaluated prediction. AutoEval-S is not given the fitted weights θ*, and the modality ablation (Table IX) shows vision is dominant (RG SRCC drops from 0.84 to 0.54 without visual context, but only to 0.81 without the physics prompt), so the RG result is not statistically forced by the kinematic summary alone. EG and CoT labels are independent human annotations; training and evaluating on them is standard supervised learning. The only self-citation is MP1 [14] (four authors overlap), used as a related-work example of 3D point-cloud policies, not as load-bearing support for any claim or uniqueness argument. The abstract promises 'Analyses of expert consistency' but the main text defers them to the Supplementary Material; missing inter-rater reliability is a ground-truth validity concern, not a circularity. Score 2 reflects the single minor, non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claims rest on the reliability of human quality annotations, the validity of kinematic smoothness/safety/efficiency metrics as proxies for those annotations, the accuracy of the policy-versus-teleoperation source labels, and the effectiveness of VLM/GRPO training. None of these are independently established within the main text beyond the reported experiments: inter-rater reliability is deferred, the GA weight-fitting uses the annotations themselves, and source labels are assumed accurate. No new physical or ontological entities are introduced; Eval-Actions and AutoEval are methodological artifacts, not invented entities.

free parameters (5)
  • GA weight vector θ = not reported
    θ = {w_vel, ..., w_len, λ_coll, λ_fail} in Eq. 1 is optimized via Genetic Algorithm to align S_raw with human rankings (Eq. 2). These weights define the RG ground-truth labels.
  • Gaussian kernel width σ = not reported
    σ in Eq. 8 controls the soft regression reward for score prediction in GRPO; it is a hand-tuned sensitivity hyperparameter.
  • Format reward weight γ = 0.2
    γ in Eq. 10 balances content accuracy and format rewards; ablated in Table VII with 0.2 chosen as optimal.
  • Task weights ω_score:ω_succ:ω_src = 4:3:3
    Weights in Eq. 9 for score, success, and source rewards; ablated in Table VIII with 4:3:3 chosen as optimal.
  • KL coefficient β = not reported
    β in Eq. 12 controls KL regularization strength in GRPO; its value is not specified.
axioms (4)
  • domain assumption Human expert annotations (EG, RG rankings, CoT reasoning) are reliable and unbiased ground truth for 'fine-grained action quality'.
    Section III-B defines the annotation protocol; inter-rater reliability is deferred to supplementary and not established in the main text.
  • domain assumption Kinematic metrics (variance of joint angular velocity/acceleration, collision/failure penalties) capture the human notion of smoothness, safety, and efficiency as used in RG label construction.
    Eq. 1 and Section IV-A assume these metrics are meaningful proxies for human quality judgments; no independent validation of this mapping is provided.
  • domain assumption The source labels (policy-generated vs teleoperated) in the dataset are accurate.
    Section III-A states trajectories come from 20 human operators and multiple VA/VLA policies, but the paper provides no audit of label correctness.
  • domain assumption A pretrained vision-language model fine-tuned with LoRA and optimized with GRPO is an appropriate and sufficient model class for action quality assessment.
    Section IV-B selects VLM + GRPO without a formal comparison to other model classes (e.g., pure kinematic regressors); the w/o Visual Context ablation (0.54 SRCC) shows kinematics alone are insufficient, supporting the model choice but not establishing optimality.

pith-pipeline@v1.3.0-alltime-deepseek · 22276 in / 16245 out tokens · 178161 ms · 2026-08-03T07:54:41.643870+00:00 · methodology

0 comments
read the original abstract

Although Vision--Action (VA) and Vision--Language--Action (VLA) policies have advanced robotic manipulation, their evaluation remains dominated by binary success rates, which obscure process-level differences among executions that complete the same task. We introduce Eval-Actions, a diagnostic evaluation methodology and real-robot benchmark for fine-grained execution-quality assessment of learned manipulation policies. Eval-Actions combines criteria-based Expert Grading (EG), Rank-Guided (RG) labels that align measurable motion indicators with expert rankings, and Chain-of-Thought-style (CoT) annotations that explain observable quality differences. The benchmark contains 13K+ teleoperated and policy-generated real-robot episodes covering 150+ tasks and approximately 52 hours of recordings with RGB-D videos, robot-state trajectories, task descriptions, and success/failure labels. Its densely annotated subset provides EG/RG/CoT supervision for training and evaluation. We further provide AutoEval, a reference multimodal evaluator that predicts quality scores, task outcomes, and diagnostic explanations from RGB temporal evidence and compact kinematic summaries. On the annotated Eval-Actions test split, AutoEval-S achieves Spearman rank correlations (SRCCs) of 0.81 and 0.84 under EG and RG, with success detection accuracies of 90.6% and 91.0%; AutoEval-P reaches 0.70 SRCC under CoT. Analyses of expert consistency, physical-metric baselines, modality ablations, structured generalization, and offline policy ranking show that Eval-Actions provides standardized, interpretable diagnostic signals complementary to success-rate evaluation.

Figures

Figures reproduced from arXiv: 2601.18723 by Hong Liu, Juyi Sheng, Mengyuan Liu, Peiming Li, Tianming Xu, Tiantian Xu, Ziyi Wang.

Figure 1
Figure 1. Figure 1: The Crisis of Evaluation Credibility and the Proposed Trust￾worthy Evaluation Solution. (Top) The Crisis: We identify two critical sources of ambiguity obstructing trustworthy evaluation: Gap 1 (Ambiguity in Execution Quality), where binary metrics mask shaky or unsafe execution (visualized as “Jerky Success” vs. “Smooth Success”), and Gap 2 (Ambiguity in Source Authenticity), where the provenance of “succ… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Eval-Actions Benchmark. The figure visualizes the dataset structure: (Left) Task Diversity: Representative snapshots from our 150+ scenarios, covering both single-arm interactions (e.g., “Throw away trash”) and complex bimanual coordination (e.g., “Tidy medicine box”). (Middle) Detailed Case Study: A specific instantiation of the “Throw away trash” task shown on the left. Crucially, each ta… view at source ↗
Figure 3
Figure 3. Figure 3: Representative Task Statistics of the Eval-Actions. The top chart [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of Expert Grading across the Eval-Actions Small subset. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the proposed AutoEval framework. The system processes a robot manipulation video sequence (e.g., 32 frames) alongside kinematic prompts. Top (AutoEval-S): Designed for Expert Grading and Rank-Guided tasks, this branch employs a Spatio-Temporal Aggregation Strategy to compress high-frequency motion details into composite visual tokens. It generates structured text predictions; following format d… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative Comparison of Fine-Grained Action Quality Assessment. The figure illustrates performance across two distinct tasks: bowl stacking and towel folding with handover. (Top) This section highlights the efficacy of the Spatio-Temporal Aggregation Strategy in AutoEval-S, which successfully captures subtle motion defects (e.g., “duplicate crawling”), whereas InternVL3.5-4B fails to detect such nuances.… view at source ↗
Figure 7
Figure 7. Figure 7: Evolution of Individual Reward Components during Training. The figure visualizes the trajectories of the cumulative reward and its constituent terms (Score, Source, Success, and Accuracy) across training steps. Notably, the score reward exhibits steady convergence, significantly contributing to the maximization of the overall objective and validating the alignment of reasoning with fine-grained quality sta… 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

45 extracted references · 16 linked inside Pith

  1. [1]

    Mahi” Shafiullah, D. Xu, K. Fragkiadaki, G. Chalvatzaki, and J. Peters, “A survey on deep generative models for robot learning from multimodal demonstrations,

    J. Urain, A. Mandlekar, Y . Du, N. Muhammad “Mahi” Shafiullah, D. Xu, K. Fragkiadaki, G. Chalvatzaki, and J. Peters, “A survey on deep generative models for robot learning from multimodal demonstrations,” IEEE Transactions on Robotics, vol. 42, pp. 60–79, 2026

    2026

  2. [2]

    Deformation control and thrust analysis of a flexible fishtail with muscle-like actuation,

    J. Gu, J. Wang, Z. Liu, M. Tan, J. Yu, and Z. Wu, “Deformation control and thrust analysis of a flexible fishtail with muscle-like actuation,”IEEE Transactions on Robotics, vol. 41, pp. 159–179, 2025

    2025

  3. [3]

    A locust- inspired robot capable of continuous crawl–jump–gliding locomotion with optimized transitional control,

    Y . Xu, W. Zhang, L. Peng, Q. Zhou, Q. Li, and Q. Shi, “A locust- inspired robot capable of continuous crawl–jump–gliding locomotion with optimized transitional control,”IEEE Transactions on Robotics, vol. 41, pp. 220–235, 2025

    2025

  4. [4]

    Biomimetic un- derwater soft snake robot: Self-motion sensing and online gait control,

    H. Shi, Y . Meng, W. Cui, M. Rao, S. Wang, and Y . Xie, “Biomimetic un- derwater soft snake robot: Self-motion sensing and online gait control,” IEEE Transactions on Robotics, vol. 41, pp. 1193–1210, 2025

    2025

  5. [5]

    Addressing human–robot symbiosis via bilevel optimization of robotic knee prosthesis control,

    W. Liu, V . Nalam, J. Si, and H. Huang, “Addressing human–robot symbiosis via bilevel optimization of robotic knee prosthesis control,” IEEE Transactions on Robotics, vol. 42, pp. 170–183, 2026

    2026

  6. [6]

    A lower limb wearable exosuit for improved sitting, standing, and walking efficiency,

    X. Zhang, E. Tricomi, X. Ma, M. Gomez-Correa, A. Ciaramella, F. Missiroli, L. Miˇskovi´c, H. Su, and L. Masia, “A lower limb wearable exosuit for improved sitting, standing, and walking efficiency,”IEEE Transactions on Robotics, vol. 41, pp. 127–140, 2025

    2025

  7. [7]

    Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,

    Z. Fu, T. Z. Zhao, and C. Finn, “Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation,” inCon- ference on Robot Learning (CoRL), 2024

    2024

  8. [8]

    Universal manipulation interface: In-the-wild robot teach- ing without in-the-wild robots,

    C. Chi, Z. Xu, C. Pan, E. Cousineau, B. Burchfiel, S. Feng, R. Tedrake, and S. Song, “Universal manipulation interface: In-the-wild robot teach- ing without in-the-wild robots,” inProceedings of Robotics: Science and Systems (RSS), 2024

    2024

  9. [9]

    Rlbench: The robot learning benchmark & learning environment,

    S. James, Z. Ma, D. Rovick Arrojo, and A. J. Davison, “Rlbench: The robot learning benchmark & learning environment,”IEEE Robotics and Automation Letters, 2020

    2020

  10. [10]

    Open x-embodiment: Robotic learning datasets and rt-x models,

    E. Collaboration and A. O. et al., “Open x-embodiment: Robotic learning datasets and rt-x models,” 2024. [Online]. Available: https://arxiv.org/abs/2310.08864

    Pith/arXiv arXiv 2024

  11. [11]

    Diffusion policy: Visuomotor policy learning via action diffusion,

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

    2023

  12. [12]

    Diffusion policy: Visuomotor policy learning via action diffusion,

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

    2024

  13. [13]

    Hierarchical diffusion policy: Manipulation trajectory generation via contact guidance,

    D. Wang, C. Liu, F. Chang, and Y . Xu, “Hierarchical diffusion policy: Manipulation trajectory generation via contact guidance,”IEEE Trans- actions on Robotics, vol. 41, pp. 2086–2104, 2025

    2086

  14. [14]

    Mp1: Meanflow tames policy learning in 1-step for robotic manipulation,

    J. Sheng, Z. Wang, P. Li, and M. Liu, “Mp1: Meanflow tames policy learning in 1-step for robotic manipulation,”arXiv preprint arXiv:2507.10543, 2025

    arXiv 2025

  15. [15]

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

    A. Brohan and N. B. et al., “Rt-1: Robotics transformer for real-world control at scale,” inarXiv preprint arXiv:2212.06817, 2022

    Pith/arXiv arXiv 2022

  16. [16]

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

    ——, “Rt-2: Vision-language-action models transfer web knowledge to robotic control,” inarXiv preprint arXiv:2307.15818, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    Open- vla: An open-source vision-language-action model,

    M. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burch- fiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn, “Open- vla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  18. [18]

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

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky, “π 0: A vision-language-action flow model for general robot control,” 2024. [Online]. Availabl...

    Pith/arXiv arXiv 2024

  19. [19]

    Bridgedata v2: A dataset for robot learning at scale,

    H. R. Walke, K. Black, T. Z. Zhao, Q. Vuong, C. Zheng, P. Hansen- Estruch, A. W. He, V . Myers, M. J. Kim, M. Duet al., “Bridgedata v2: A dataset for robot learning at scale,” inConference on Robot Learning. PMLR, 2023, pp. 1723–1736

    2023

  20. [20]

    On bringing robots home,

    N. M. M. Shafiullah, A. Rai, H. Etukuru, Y . Liu, I. Misra, S. Chin- tala, and L. Pinto, “On bringing robots home,”arXiv preprint arXiv:2311.16098, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot,

    H.-S. Fang, H. Fang, Z. Tang, J. Liu, C. Wang, J. Wang, H. Zhu, and C. Lu, “Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 653–660

    2024

  22. [22]

    Droid: A large-scale in-the-wild robot manipulation dataset,

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karam- cheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Elliset al., “Droid: A large-scale in-the-wild robot manipulation dataset,” inProceedings of Robotics: Science and Systems (RSS), 2024

    2024

  23. [23]

    Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation,

    K. Wu, C. Hou, J. Liu, Z. Che, X. Ju, Z. Yang, M. Li, Y . Zhao, Z. Xu, G. Yanget al., “Robomind: Benchmark on multi-embodiment intelligence normative data for robot manipulation,” inProceedings of Robotics: Science and Systems (RSS), 2025

    2025

  24. [24]

    Bridge data: Boosting generaliza- tion of robotic skills with cross-domain datasets,

    F. Ebert, Y . Yang, K. Schmeckpeper, B. Bucher, G. Georgakis, K. Dani- ilidis, C. Finn, and S. Levine, “Bridge data: Boosting generaliza- tion of robotic skills with cross-domain datasets,”arXiv preprint arXiv:2109.13396, 2021

    Pith/arXiv arXiv 2021

  25. [25]

    Learning hand-eye coordination for robotic grasping with deep learning and large- scale data collection,

    S. Levine, P. Pastor, A. Krizhevsky, J. Ibarz, and D. Quillen, “Learning hand-eye coordination for robotic grasping with deep learning and large- scale data collection,”The International journal of robotics research, vol. 37, no. 4-5, pp. 421–436, 2018

    2018

  26. [26]

    Grasp like humans: Learning generalizable multifingered grasping from human proprioceptive sensorimotor integration,

    C. Guo, X. Chen, Z. Zeng, Z. Guo, Y . Li, H. Xiao, D. Hu, and H. Lu, “Grasp like humans: Learning generalizable multifingered grasping from human proprioceptive sensorimotor integration,”IEEE Transactions on Robotics, vol. 41, pp. 5700–5719, 2025

    2025

  27. [27]

    Shear-based grasp control for multifingered underactuated tactile robotic hands,

    C. J. Ford, H. Li, M. G. Catalano, M. Bianchi, E. Psomopoulou, and N. F. Lepora, “Shear-based grasp control for multifingered underactuated tactile robotic hands,”IEEE Transactions on Robotics, vol. 41, pp. 3113– 3128, 2025

    2025

  28. [28]

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

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,”arXiv preprint arXiv:2403.03954, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Perceiver-actor: A multi-task transformer for robotic manipulation,

    M. Shridhar, L. Manuelli, and D. Fox, “Perceiver-actor: A multi-task transformer for robotic manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 785–799

    2023

  30. [30]

    Rvt: Robotic view transformer for 3d object manipulation,

    A. Goyal, J. Xu, Y . Guo, V . Blukis, Y .-W. Chao, and D. Fox, “Rvt: Robotic view transformer for 3d object manipulation,” inConference on Robot Learning. PMLR, 2023, pp. 694–710

    2023

  31. [31]

    Octo: An open-source generalist robot policy,

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xuet al., “Octo: An open-source generalist robot policy,”arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azharet al., “Llama: Open and efficient foundation language models,”arXiv preprint arXiv:2302.13971, 2023

    Pith/arXiv arXiv 2023

  33. [33]

    Bilateral interaction for local-global collaborative perception in low-light image enhancement,

    R. Xu, Y . Li, Y . Niu, H. Xu, Y . Chen, and T. Zhao, “Bilateral interaction for local-global collaborative perception in low-light image enhancement,”IEEE Transactions on Multimedia, vol. 26, pp. 10 792– 10 804, 2024

    2024

  34. [34]

    Human-centric fine-grained action qual- ity assessment,

    J. Xu, S. Yin, and Y . Peng, “Human-centric fine-grained action qual- ity assessment,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 8, pp. 6242–6255, 2025

    2025

  35. [35]

    Phi: Bridging domain shift in long-term action quality assessment via progressive hierarchical instruction,

    K. Zhou, H. P. Shum, F. W. Li, X. Zhang, and X. Liang, “Phi: Bridging domain shift in long-term action quality assessment via progressive hierarchical instruction,”IEEE Transactions on Image Processing, 2025

    2025

  36. [36]

    Basket: A large-scale video dataset for fine-grained skill estimation,

    Y . Pan, C. Zhang, and G. Bertasius, “Basket: A large-scale video dataset for fine-grained skill estimation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 28 952–28 962

    2025

  37. [37]

    Towards unified surgical skill assessment,

    D. Liu, Q. Li, T. Jiang, Y . Wang, R. Miao, F. Shan, and Z. Li, “Towards unified surgical skill assessment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 9522– 9531

    2021

  38. [38]

    Sedskill: Surgical events driven method for skill assessment from thoracoscopic surgical videos,

    X. Ding, X. Xu, and X. Li, “Sedskill: Surgical events driven method for skill assessment from thoracoscopic surgical videos,” inSpringer-Verlag, 2023, p. 35–45. [Online]. Available: https: //doi.org/10.1007/978-3-031-43996-4 4

    doi:10.1007/978-3-031-43996-4 2023

  39. [39]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models,

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wuet al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,”arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Biet al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,”arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    Internvl3. 5: Advancing open-source multi- modal models in versatility, reasoning, and efficiency,

    W. Wang, Z. Gao, L. Gu, H. Pu, L. Cui, X. Wei, Z. Liu, L. Jing, S. Ye, J. Shaoet al., “Internvl3. 5: Advancing open-source multi- modal models in versatility, reasoning, and efficiency,”arXiv preprint arXiv:2508.18265, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Qwen3 technical report,

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lvet al., “Qwen3 technical report,”arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  43. [43]

    Smolvlm: REGULAR PAPER 14 Redefining small and efficient multimodal models,

    A. Marafioti, O. Zohar, M. Farr ´e, M. Noyan, E. Bakouch, P. Cuenca, C. Zakka, L. B. Allal, A. Lozhkov, N. Taziet al., “Smolvlm: REGULAR PAPER 14 Redefining small and efficient multimodal models,”arXiv preprint arXiv:2504.05299, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Qwen2. 5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tanget al., “Qwen2. 5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.” ICLR, vol. 1, no. 2, p. 3, 2022

    2022