Robotics · arXiv:2609.01596 v1 · submitted 1 Sep 2026 · 54 references
Facet-0: A Robotic Foundation Model for Contact-Rich Precise Manipulation
The paper says
“Real-world robotic assembly at sub-millimeter tolerances demands spatial precision, compliant interaction, and robustness to contact failures. We present Facet-0, a robotic foundation model that predicts and values the contact consequences of its actions.”
more of the abstract
Facet-0 unifies multimodal representation learning and reinforcement learning (RL) post-training around a joint action-wrench proposal: a causal wrench history is aligned with vision-language semantics and kinematic state, and flow matching generates each action chunk together with the future wrist-wrench profile it is expected to induce. Deployment rollouts train a distributional Action-Wrench Critic to distinguish motions with similar task progress but different contact outcomes, while phase-aware rewards and contact-selective credit concentrate policy improvement on decisive interactions. To accommodate part-specific dynamics, a lightweight bounded actor reuses the frozen representation for on-robot adaptation; RL remains defined over executable Cartesian actions, while an auxiliary wrench head preserves predictive, non-commanded action-contact coupling. Trained on ManuFacet-1K, a 1,000-hour force-synchronized corpus spanning three embodiments and multiple manufacturing cells, the bounded task-adapted system reaches 82% mean success on five sub-millimeter computer-assembly tasks, compared with 15% for the strongest baseline, with 0.5 mm placement accuracy and 50 ms command latency.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 8 Sep 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Impressive system, real novelty in joint action-wrench prediction and value-guided refinement, but the unvalidated VLM judge Ω that generates every RL label is a load-bearing gap that prevents full acceptance. This line is wrong
3 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 54 references say.
The paper says, at Section 5.1, Eq. (4)
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The VLM judge Ω is the sole source of sub-goal completion labels, contact-phase labels φ_t, and per-phase force envelopes used in the reward r_t. Every downstream RL signal — the temporal-difference target in Eq. (5), the credit score δ^{(N)}_t in Eq.
more of objection 1
(6), and the positive-interaction label b_t in Eq. (7) — originates from Ω. No held-out accuracy, precision/recall, per-phase confusion matrix, or annotator agreement is reported for Ω. If Ω mislabels sub-goal completion under visually ambiguous contact (e.g., jamming vs. clean insertion) or misassigns contact phases, the reward violation term and all subsequent credit are miscalibrated. The controlled 16%→38%→82% progression and the mechanism claim that 'contact therefore influences which interactions the policy improves, not only what it observes' (Sec. 2.2) rest on these labels being correct. The authors should report validation of Ω on a held-out set of annotated rollouts, including per-phase accuracy and agreement with human annotators, and show how errors in Ω would affect the measured success progression.
The paper says, at Section 6.1 and Appendix A.1
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The strongest baseline (π0.5+RECAP-style) against which the headline 15% success is compared is a reimplementation of the RECAP advantage-conditioning objective [41] on the π0.5 backbone, not the released π★_0.6 model or a published baseline. The paper does not compare the performance of this reimplementation against the original RECAP results on equivalent tasks.
more of objection 2
Because the 15% figure defines the headline margin, any deviation between the reimplementation and the original RECAP behavior (e.g., due to differences in reward design, advantage estimation, or hyperparameters) could affect the reported improvement ratio. The authors should document how closely their reimplementation reproduces the published RECAP algorithm and, if possible, provide a direct comparison on a common set of tasks.
The paper says, at Table 4 and Section 6.5
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The contact-role ablation on RAM insertion uses 20 trials per variant. The paper correctly notes that individual removals (predictive wrench, critic wrench, adaptation wrench) differ by only one to two successes and that 'their ordering is not resolved.' However, the takeaway states that 'removing every predictive and evaluative role of wrench is the one degradation that is large relative to the measurement resolution.'
more of objection 3
Even the joint removal (w/o all contact roles) shows a reduction from 95% to 45% — a 50-point drop over 20 trials. While the 20-trial resolution cannot resolve small differences, the joint removal result should be accompanied by a confidence interval or a statement of statistical significance to support the claim that the degradation is large relative to measurement resolution.
6 minor objections
The paper says, at Abstract
Pith says: minor objection 1 of 6
The abstract states '0.5 mm placement accuracy' but the rest of the paper uses '0.5mm' without a space. Please be consistent.
The paper says, at Section 3.1
Pith says: minor objection 2 of 6
The British spelling 'synchronised' appears in the first paragraph of Section 3.1, while 'synchronized' is used elsewhere. Please use a single spelling convention.
show 4 more
The paper says, at Table 2 and Section 6.1
Pith says: minor objection 3 of 6
The controlled variant 'Align' in Table 2 is introduced without a full definition; its description in Section 6.1 ('learns the semantic–contact representation') is clear only after reading Section 4. Please add a brief one-sentence definition in the table caption or in the paragraph that introduces the variants.
The paper says, at Figure 2 caption
Pith says: minor objection 4 of 6
The caption says 'at 0.5mm action resolution' but the abstract and elsewhere use '0.5 mm'. Please make the spacing consistent throughout.
The paper says, at Appendix D.2
Pith says: minor objection 5 of 6
In the description of contact intensity c_t, the wording 'causally smoothed wrench differences' is vague. Please clarify the smoothing window length and the method (e.g., exponential moving average with a specific α).
The paper says, at References
Pith says: minor objection 6 of 6
Reference [29] (Lipman et al., Flow Matching for Generative Modeling) appears twice in the reference list (also at [6]? Actually [6] is PaliGemma, that's fine). The reference list seems complete, but please check that all citations in the text have corresponding entries.
What would settle it
Using the stored deployment rollouts from Section 6.2, hold out roughly 200 phase-segmented episodes and have independent human annotators label Ω_t and φ_t. Compute Ω's precision/recall and per-phase confusion for completions and contact phases, plus inter-annotator agreement on a subset. Then retrain the critic and rerun the +RL and Full variants on the five-task suite with rewards relabeled by the human annotations instead of Ω_t/φ_t, holding all other settings and seeds fixed. If the Full mean drops by more than about 20 points or the align/place sub-goal rows shift by more than about 10 points, the unvalidated judge is generating part of the headline; if performance is unchanged, the concern is resolved. This check requires releasing the rollouts, Ω weights, and training recipe, none of which currently appear in the paper.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 8 Sep 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The headline result—82% vs 15% on sub-millimeter assembly, with a clean 16→38→82 progression—is striking, and the arithmetic is consistent throughout. The joint action-wrench proposal is genuinely novel: no prior VLA paper treats anticipated wrench as a decoding target and then scores that joint proposal with a distributional critic gated by contact regime. The ablation design is honest (they admit individual removals don't establish a ranking), and the paper is unusually transparent about its limitations (Fig. 10 labeled conceptual, Sec. C disclaims certified force bounds). I believe the system works.
That said, the VLM judge Ω is the single weakest point, and the stress-test note is right. Every reward, TD target, credit score, and force envelope comes from Ω's sub-goal completion and contact-phase labels. Section 5.1 trains Ω but reports zero accuracy, precision/recall, or annotator agreement. If Ω mislabels a visually similar jam as clean completion, the entire 16→38→82 progression and the clean-vs-jam separation in Fig. 4 become properties of the labeler, not of the predicted wrench. This is not a minor oversight—it is load-bearing, and it is testable. The paper needs an offline validation of Ω against human labels, ideally per-phase confusion matrices. Without that, the mechanism claim that "contact changes which interactions the policy improves" is not established, no matter how clean the internal math.
Other soft spots: the strongest baseline (π0.5+RECAP) is an internal reimplementation with no release to verify parity; the 20-trial cells lack error bars or confidence intervals; no code, data, or checkpoints are provided. The reward's force-violation term also overlaps with the success metric, encoding some of the measured contact quality into the objective. These are real but secondary.
Who this is for: anyone working on VLA post-training or force-aware manipulation. The paper deserves a serious referee: the core ideas are novel and the empirical picture, though incomplete, is coherent. But it needs major revisions: validate Ω, release code/data, report error bars, and clarify baseline fidelity.
Recommendation: accept with major revisions, contingent on the VLM validation and a code release.
The strongest objection
Every reward, temporal-difference target, credit score, and force-envelope assignment in the value-guided pipeline originates from the VLM judge Ω. Section 5.1 trains Ω to segment each rollout into sub-goals, marking completion with Ω_t in {0,1} and labeling the contact phase φ_t; Eq. (4) converts those labels into r_t, Eq. (5) trains the Action–Wrench Critic from r_t, Eq. (6) computes the credit score δ_t from V and r, and Eq. (7) uses the resulting label b_t and contact intensity c_t to select refinement samples. No held-out accuracy, precision/recall, per-phase confusion, or annotator agreement is reported for Ω, although Section 3.2 already relies on the same VLM pass to verify phase boundaries. If Ω mislabels sub-goal completion under jamming or confuses contact phases, the per-phase force envelope used in Eq. (4) and all downstream credit are miscalibrated; the Fig. 4 separation of a clean insertion from a visually similar jam would then be a property of the labeler rather than of the predicted wrench. This is load-bearing because the headline rests on the controlled 16 to 38 to 82 progression and on the mechanism claim that contact changes which interactions the policy improves. The concern is amplified by the reward's violation term and the success metric both keying off per-phase force limits, so part of the measured contact-quality gain is encoded in the objective; even that encoding is gated by φ_t from Ω. The issue is concrete and testable, and resolving it does not require assuming the results are fabricated.
The reader's weakest assumption already identified Ω as the first load-bearing premise, and the stress-test agrees with that identification. The paper's arithmetic and internal consistency checks pass, and the self-asserted limitations about ablation ordering, force bounds, and conceptual figures lower soundness without breaking the headline by themselves. But the RL half of the paper cannot be checked from the text because every reward label is produced by an instrument with no reported validation. The concern is not that the numbers are impossible; the 16→38→82 progression could be real if Ω is accurate. It is that the central mechanism claim and the value-guided gains are conditioned on an unvalidated labeler, and the paper provides no evidence to break that conditioning. Baseline matching, the absence of released code and data, and same-fixture evaluation are additional reasons not to accept, but they are secondary to Ω because even a perfectly matched baseline leaves the reported gains unexplained if the reward labels are unreliable. The reader's CONDITIONAL verdict therefore stands unchanged; I would not move it to ACCEPT without validating Ω, nor to REJECT without evidence that the labels are actually wrong. The concrete test above is the smallest experiment that would settle the matter.
The referee's report
If the claims hold, Facet-0 makes a strong step toward bridging generalist vision–language–action policies with the precision demands of contact-rich manufacturing. The joint action–wrench prediction, the use of an Action–Wrench Critic that scores the predicted contact consequence, and the bounded local adaptation mechanism are well-motivated and differ from prior work that treats force only as an input or feedback signal. The ManuFacet-1K dataset contributes a large, force-synchronized resource that may benefit the broader manipulation community. The empirical results are quantitatively impressive: the headline 82% task success with 0.5 mm placement accuracy and 50 ms latency, the systematic ablation from 16% to 82%, and the reduction in human intervention from 47% to 24% are all noteworthy. However, the evidence for these results depends critically on the reliability of the VLM judge that provides the reward, credit, and force-envelope labels; this dependency is not yet validated in the manuscript.
Pith's reading of the paper (a machine's summary, not the authors')
Facet-0 is a robotic foundation model that treats contact as a predicted consequence of its proposed actions and uses deployment experience to value those consequences. Its core innovation is joint action–wrench flow matching: the model generates both the Cartesian motion command and the six-axis wrist wrench it is expected to produce, learned from a 1,000-hour force-synchronized dataset. A distributional Action–Wrench Critic then scores this joint proposal, allowing the system to distinguish motions with similar geometric progress but different contact outcomes (e.g., clean insertion vs. jamming). On five sub-millimeter computer-assembly tasks, the full system reaches 82% mean success, compared with 15% for the strongest baseline, with 0.5 mm placement accuracy and 50 ms command latency. The results demonstrate that anticipating and valuing contact is necessary and sufficient for reliable precision assembly at these tolerances.
What carries the argument. The central mechanism is the joint action–wrench proposal generated by conditional flow matching through a semantic–contact representation. The action chunk and its anticipated wrist wrench are decoded together from a fused representation of vision–language, kinematic state, and causal wrench history. This joint proposal is then scored by a distributional Action–Wrench Critic trained on deployment rollouts, allowing the system to separate motions with similar geometric progress but different contact outcomes. Contact-selective credit and bounded local adaptation further refine the policy toward decisive interactions.
Core claim. Facet-0 establishes that a generalist robot policy can anticipate and value the contact consequences of its actions, and that doing so dramatically improves reliability in sub-millimeter assembly. The core discovery is that joint action–wrench flow matching (generating both the planned Cartesian command and the six-axis wrench it will produce) creates a representation where the policy can learn which contact outcomes are desirable. This representation, combined with a distributional Action–Wrench Critic that evaluates the joint proposal and a contact-selective credit assignment mechanism, enables policy improvement to focus on decisive interactions that determine task success. The empirical result—82% mean success across five tasks vs. 15% for the strongest baseline, with 0.5 mm accuracy—supports the claim that predicting and valuing contact is necessary and sufficient for precision assembly at these tolerances.
The premise it rests on. Every reward, critic target, and credit label depends on the VLM judge Ω that segments rollouts and labels contact phases; the paper reports no accuracy or precision/recall for this judge, so if Ω systematically mislabels contact outcomes, the entire value-guided improvement rests on unreliable labels.
If the paper is right (Pith's inferences, not the authors')
Automatic recovery from contact failures becomes feasible because wrench-based signals allow the system to detect and correct misalignments that vision alone cannot distinguish.
The frozen semantic–contact representation supports local adaptation to unseen parts: with only ten demonstrations and 6.6% of parameters updated, the system reaches 45% success on an unseen memory module, compared with 5% for baselines.
Value-guided post-training from deployment experience outperforms advantage-weighted behavior cloning: Facet-0 reduces human interventions from 47% to 24% and increases failure recovery from 44% to 81% on the Disk task.
The 82% mean success across five tasks and 87% mean on 13 contact-critical sub-goals indicates that the representation and adaptation jointly address the contact bottleneck that limits generalist VLA policies in precision domains.
Deployment properties (50 ms latency, 0.5 mm accuracy) make the model suitable for real-time factory use.
If the VLM judge used for reward annotation has a failure mode on specific contact geometries (e.g., glossy or specular parts), the entire value-guided improvement chain may be miscalibrated; independent evaluation of the judge's accuracy on the full contact-phase taxonomy would strengthen the claim.
The method depends on access to wrist-mounted six-axis force/torque sensing. Extending the approach to cells without F/T sensors, perhaps through force-distilled vision-only models, would determine whether the gains are sensor-specific or can be achieved with proxy signals.
The 82% success rate on LEVER (the only purely contact-critical task) suggests that purely contact-critical chains remain the hardest; this points to a need for better exploration or credit assignment in fully contact-constrained settings.
The bounded local adaptation paradigm, which reuses a frozen representation and policy reference, could be applied to other domains where local dynamics change but task semantics remain stable, such as garment manipulation or medical instrument handling.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Real-world robotic assembly at sub-millimeter tolerances demands spatial precision, compliant interaction, and robustness to contact failures. We present Facet-0, a robotic foundation model that predicts and values the contact consequences of its actions. Facet-0 unifies multimodal representation learning and reinforcement learning (RL) post-training around a joint action-wrench proposal: a causal wrench history is aligned with vision-language semantics and kinematic state, and flow matching generates each action chunk together with the future wrist-wrench profile it is expected to induce. Deployment rollouts train a distributional Action-Wrench Critic to distinguish motions with similar task progress but different contact outcomes, while phase-aware rewards and contact-selective credit concentrate policy improvement on decisive interactions. To accommodate part-specific dynamics, a lightweight bounded actor reuses the frozen representation for on-robot adaptation; RL remains defined over executable Cartesian actions, while an auxiliary wrench head preserves predictive, non-commanded action-contact coupling. Trained on ManuFacet-1K, a 1,000-hour force-synchronized corpus spanning three embodiments and multiple manufacturing cells, the bounded task-adapted system reaches 82% mean success on five sub-millimeter computer-assembly tasks, compared with 15% for the strongest baseline, with 0.5 mm placement accuracy and 50 ms command latency.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23
The authors' main claim, as the classifier picked it out 0.97
“Trained on ManuFacet-1K, a 1,000-hour force-synchronized corpus spanning three embodiments and multiple manufacturing cells, the bounded task-adapted system reaches 82% mean success on five sub-millimeter computer-assembly tasks, compared with 15% for the strongest baseline, with 0.5 mm placement accuracy and 50 ms command latency.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.16 · human participants 0.04 · controlled comparison 0.97 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.04
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Real-world robotic assembly at sub-millimeter tolerances demands spatial precision, compliant interaction, and robustness to contact failures. | background 1.00 | 0.05 | 0.03 | asserted 1.00 |
| We present Facet-0, a robotic foundation model that predicts and values the contact consequences of its actions. | contribution 1.00 | 0.88 | 0.21 | asserted 1.00 |
| Facet-0 unifies multimodal representation learning and reinforcement learning (RL) post-training around a joint action-wrench proposal: a causal wrench history is aligned with vision-language semantics and kinematic state, and flow matching generates each action chunk together with the future wrist-wrench profile it is expected to induce. | method 0.95 | 0.86 | 0.09 | asserted 1.00 |
| Deployment rollouts train a distributional Action-Wrench Critic to distinguish motions with similar task progress but different contact outcomes, while phase-aware rewards and contact-selective credit concentrate policy improvement on decisive interactions. | method 0.99 | 0.88 | 0.07 | asserted 0.98 |
| To accommodate part-specific dynamics, a lightweight bounded actor reuses the frozen representation for on-robot adaptation; RL remains defined over executable Cartesian actions, while an auxiliary wrench head preserves predictive, non-commanded action-contact coupling. | method 0.99 | 0.85 | 0.05 | asserted 1.00 |
| Trained on ManuFacet-1K, a 1,000-hour force-synchronized corpus spanning three embodiments and multiple manufacturing cells, the bounded task-adapted system reaches 82% mean success on five sub-millimeter computer-assembly tasks, compared with 15% for the strongest baseline, with 0.5 mm placement accuracy and 50 ms command latency. | result 1.00 | 0.96 | 0.66 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 1.00
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 5,210 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
References (54)
- [41] Physical Intelligence.𝜋★ 0.6: A vla that learns from experience.Technical report, 2025 on Pith
- [1] Fares J. Abu-Dakka and Matteo Saveriano. Variable impedance control and learning: A review.Frontiers in Robotics and AI, 7:590681, 2020 on Pith
- [2] AgiBot-World-Contributors, Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao,XindongHe,etal. Agibotworldcolosseo: Alarge-scalemanipulationplatformforscalableandintelligent embodied systems.arXiv preprint arXiv:2503.06669, 2025 on PithPith's reviewarXiv
- [3] Lars Ankile, Zhenyu Jiang, Rocky Duan, Guanya Shi, Pieter Abbeel, and Anusha Nagabandi. Residual off-policy rl for finetuning behavior cloning policies.arXiv preprint arXiv:2509.19301, 2025 on PitharXiv
- [4] Marc G. Bellemare, Will Dabney, and Rémi Munos. A distributional perspective on reinforcement learning. In International Conference on Machine Learning (ICML), 2017 on Pith
- [5] Cristian C. Beltran-Hernandez, Damien Petit, Ixchel G. Ramirez-Alpizar, Takayuki Nishi, Shinichi Kikuchi, Takamitsu Matsubara, and Kensuke Harada. Learning force control for contact-rich manipulation tasks with rigid position-controlled robots.IEEE Robotics and Automation Lett... on Pith
- [6] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer.arXiv preprint arXiv:2407.07726, 2024 on PithPith's reviewarXiv
- [7] Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Robert Equi, ChelseaFinn,NiccoloFusai,etal. 𝜋0.5: Avision-language-actionmodelwithopen-worldgeneralization. InProceedings of the 9th Conference on Robot Learning, volume 305 ofProceed... on Pith
- [8] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.𝜋0: A vision-language-action flow model for general robot control. InProceedings of Robotics: Science and Systems, 2025 on Pith
- [9] Yuhui Chen, Shuai Tian, Shugao Liu, Yingting Zhou, Haoran Li, and Dongbin Zhao. ConRFT: A reinforced fine-tuning method for VLA models via consistency policy. InProceedings of Robotics: Science and Systems, 2025 on Pith
- [10] Cheng Chi, Siyuan Feng, Yilun Du, Zhenjia Xu, Eric Cousineau, Benjamin Burchfiel, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. InRobotics: Science and Systems (RSS), 2023 on Pith
- [11] Haoyuan Deng, Yitong Gao, Yudong Lin, Haichao Liu, Zhenyu Wu, and Ziwei Wang. UniIntervene: Agentic intervention for efficient real-world reinforcement learning.arXiv preprint arXiv:2606.12372, 2026 on PithPith's reviewarXiv
- [12] Haoyuan Deng, Yudong Lin, Yuanjiang Xue, Haoyang Du, Qianzhun Wang, Boyang Zhou, Zhenyu Wu, and ZiweiWang. E2HiL:Entropy-guidedsampleselectionforefficientreal-worldhuman-in-the-loopreinforcement learning.IEEE Robotics and Automation Letters, 11(7):8084–8091, 2026 on Pith
- [13] Haoyuan Deng, Zhenyu Wu, Haichao Liu, Wenkai Guo, Yuquan Xue, Ziyu Shan, Chuanrui Zhang, Bofang Jia, Yuan Ling, Guanxing Lu, et al. A survey on reinforcement learning of vision-language-action models for robotic manipulation.TechRxiv, 2025 on Pith
- [14] Danny Driess, Jost Tobias Springenberg, Brian Ichter, Lili Yu, Adrian Li-Bell, Karl Pertsch, Allen Z. Ren, Homer Walke, Quan Vuong, Lucy Xiaoyang Shi, and Sergey Levine. Knowledge insulating vision-language- action models: Train fast, run fast, generalize better. InAdvances in... on Pith
- [15] Federica Ferraguti, Cristian Secchi, and Cesare Fantuzzi. A tank-based approach to impedance control with variable stiffness. InIEEE International Conference on Robotics and Automation (ICRA), 2013 on Pith
- [16] Scott Fujimoto and Shixiang Shane Gu. A minimalist approach to offline reinforcement learning. InAdvances in Neural Information Processing Systems (NeurIPS), 2021 on Pith
- [17] Scott Fujimoto, Herke van Hoof, and David Meger. Addressing function approximation error in actor-critic methods. InInternational Conference on Machine Learning (ICML), 2018 on Pith
- [18] Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Arenas,TravisArmstrong,AshwinBalakrishna,RobertBaruch,MariaBauza,etal. Geminirobotics: Bringing ai into the physical world.arXiv preprint arXiv:2503.20020, 2025 on PithPith's reviewarXiv
- [19] Zihao He, Hongjie Fang, Jingjing Chen, Hao-Shu Fang, and Cewu Lu. Foar: Force-aware reactive policy for contact-rich robotic manipulation.IEEE Robotics and Automation Letters, 10(6):5625–5632, 2025 on Pith
show 34 more
- [20] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS Workshop on Deep Generative Models and Downstream Applications, 2021. arXiv:2207.12598 on PithPith's reviewarXiv
- [21] Neville Hogan. Impedance control: An approach to manipulation, parts i–iii.ASME Journal of Dynamic Systems, Measurement, and Control, 107(1):1–24, 1985 on Pith
- [22] Jialei Huang, Shuo Wang, Fanqi Lin, Yihang Hu, Chuan Wen, and Yang Gao. Tactile-VLA: Unlocking vision-language-actionmodel’sphysicalknowledgefortactilegeneralization.arXivpreprintarXiv:2507.09160, 2025 on PithPith's reviewarXiv
- [23] Tadanobu Inoue, Giovanni De Magistris, Asim Munawar, Tsuyoshi Yokoya, and Ryuki Tachibana. Deep reinforcement learning for high precision assembly tasks. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017 on Pith
- [24] Tobias Johannink, Shikhar Bahl, Ashvin Nair, Jianlan Luo, Avinash Kumar, Matthias Loskyll, Juan Aparicio Ojea, Eugen Solowjow, and Sergey Levine. Residual reinforcement learning for robot control. InIEEE International Conference on Robotics and Automation (ICRA), 2019 on Pith
- [25] Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, et al. Droid: A large-scale in-the-wild robot manipulation dataset. InRobotics: Science and Systems (RSS), 2024 on Pith
- [26] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P. Foster, Pannag R. Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An ... on Pith
- [27] Klas Kronander and Aude Billard. Stability considerations for variable impedance control.IEEE Transactions on Robotics, 32(5):1298–1305, 2016 on Pith
- [28] Kun Lei, Huanyu Li, Dongjie Yu, Zhenyu Wei, Lingxiao Guo, Zhennan Jiang, Ziyu Wang, Shiyu Liang, and Huazhe Xu. RL-100: Performant robotic manipulation with real-world reinforcement learning.arXiv preprint arXiv:2510.14830, 2025 on PitharXiv
- [29] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations (ICLR), 2023 on Pith
- [30] Guanxing Lu, Wenkai Guo, Chubin Zhang, Yuheng Zhou, Haonan Jiang, Zifeng Gao, Yansong Tang, and Ziwei Wang. VLA-RL: Towards masterful and general robotic manipulation with scalable reinforcement learning.arXiv preprint arXiv:2505.18719, 2025 on PithPith's reviewarXiv
- [31] Jianlan Luo, Zheyuan Hu, Charles Xu, You Liang Tan, Jacob Berg, Archit Sharma, Stefan Schaal, Chelsea Finn, Abhishek Gupta, and Sergey Levine. Serl: A software suite for sample-efficient robotic reinforcement learning. InIEEE International Conference on Robotics and Automation... on Pith
- [32] Jianlan Luo, Charles Xu, Fangchen Liu, Liam Tan, Zipeng Lin, Jeffrey Wu, Pieter Abbeel, and Sergey Levine. FMB: A functional manipulation benchmark for generalizable robotic learning.International Journal of Robotics Research, 44(4):592–606, 2025 on Pith
- [33] Jianlan Luo, Charles Xu, Jeffrey Wu, and Sergey Levine. Precise and dexterous robotic manipulation via human-in-the-loop reinforcement learning.arXiv preprint arXiv:2410.21845, 2024 on PithPith's reviewarXiv
- [34] Matthew T. Mason. Compliance and force control for computer controlled manipulators.IEEE Transactions on Systems, Man, and Cybernetics, 11(6):418–432, 1981 on Pith
- [35] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. RoboCasa: Large-scale simulation of household tasks for generalist robots. In Robotics: Science and Systems, 2024 on Pith
- [36] NVIDIA. NVIDIA Isaac GR00T N1.7: A foundation model for generalist robots.https://github.com/ NVIDIA/Isaac-GR00T/releases/tag/n1.7-release, 2026. Official N1.7 release, accessed 2026-08-28 on Pith
- [37] NVIDIA, Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025 on PithPith's reviewarXiv
- [38] Open X-Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, et al. Open x-embodiment: Robotic learning datasets and rt-x models. InIEEE International Conference on Robotics and Automation (ICRA), 2024 on Pith
- [39] Xue Bin Peng, Aviral Kumar, Grace Zhang, and Sergey Levine. Advantage-weighted regression: Simple and scalable off-policy reinforcement learning.arXiv preprint arXiv:1910.00177, 2019 on PithPith's reviewarXiv
- [40] Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models. InProceedings of Robotics: Science and Systems, 2025 on Pith
- [42] Marc H. Raibert and John J. Craig. Hybrid position/force control of manipulators.ASME Journal of Dynamic Systems, Measurement, and Control, 103(2):126–133, 1981 on Pith
- [43] Tom Silver, Kelsey Allen, Josh Tenenbaum, and Leslie Kaelbling. Residual policy learning.arXiv preprint arXiv:1812.06298, 2018 on PithPith's reviewarXiv
- [44] Daniel Sliwowski, Shail Jadav, Sergej Stanovcic, Jedrzej Orbik, Johannes Heidersberger, and Dongheui Lee. REASSEMBLE: A multimodal dataset for contact-rich robotic assembly and disassembly.arXiv preprint arXiv:2502.05086, 2025 on PithPith's reviewarXiv
- [45] Markku Suomalainen, Yiannis Karayiannidis, and Ville Kyrki. A survey of robot manipulation in contact. Robotics and Autonomous Systems, 156:104224, 2022 on Pith
- [46] LuigiVillaniandJorisDeSchutter. Forcecontrol. InSpringerHandbookofRobotics,pages195–220.Springer, 2nd edition, 2016 on Pith
- [47] AndrewWagenmaker,YunchuZhang,MitsuhikoNakamoto,SeohongPark,WaleedYagoub,AnushaNagabandi, Abhishek Gupta, and Sergey Levine. Steering your diffusion policy with latent space reinforcement learning. InProceedings of the 9th Conference on Robot Learning, volume 305 ofProceedings ... on Pith
- [48] Kun Wu, Chengkai Hou, Jiaming Liu, Zhengping Che, Xiaozhu Ju, Zhuqin Yang, Meng Li, Yinuo Zhao, Zhiyuan Xu, Guang Yang, et al. RoboMIND: Benchmark on multi-embodiment intelligence normative data for robot manipulation. InProceedings of Robotics: Science and Systems, 2025 on Pith
- [49] Jiawen Yu, Hairuo Liu, Qiaojun Yu, Jieji Ren, Ce Hao, Haitong Ding, Guangyu Huang, Guofan Huang, Yan Song,PanpanCai,CewuLu,andWenqiangZhang.ForceVLA:EnhancingVLAmodelswithaforce-awareMoE for contact-rich manipulation.Advances in Neural Information Processing Systems, 2025. arX... on PitharXiv
- [50] Xiu Yuan, Tongzhou Mu, Stone Tao, Yunhao Fang, Mengke Zhang, and Hao Su. Policy decorator: Model- agnostic online refinement for large policy model. InInternational Conference on Learning Representations, 2025 on Pith
- [51] Shuoheng Zhang, Yifu Yuan, Hongyao Tang, Yan Zheng, Qiaojun Yu, Pengyi Li, Guowei Huang, Helong Huang, Xingyue Quan, and Jianye Hao. ForceFlow: Learning to feel and act via contact-driven flow matching. arXiv preprint arXiv:2605.11048, 2026 on PithPith's reviewarXiv
- [52] Zongzheng Zhang, Haobo Xu, Zhuo Yang, Chenghao Yue, Zehao Lin, Huan-ang Gao, Ziwei Wang, and Hao Zhao. Elucidating the design space of torque-aware vision-language-action models. InProceedings of the 9th Conference on Robot Learning, volume 305 ofProceedings of Machine Learnin... on Pith
- [53] Ruiteng Zhao, Wenshuo Wang, Yicheng Ma, Xiaocong Li, Francis E. H. Tay, Marcelo H. Ang, and Haiyue Zhu. FD-VLA: Force-distilled vision-language-action model for contact-rich manipulation.arXiv preprint arXiv:2602.02142, 2026. 18 A Assembly Suite Details The suite uses a shared... on PitharXiv
- [54] trained on all rollouts under progress-calibrated targets. The concrete distributional parameterization and regression distance are implementation-dependent; the method requires only a return distribution whose mean defines𝑄𝜓. Four auxiliary heads share the critic trunk and ar... on Pith
Cited by 3 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- CW-WAM: A Structured Contact-Wrench World Action Model for Force-Sensitive Manipulation cs.RO · Oct 2026
- ForceBench: Benchmarking Force Integration for Robotic Manipulation cs.RO · Oct 2026
- UniIntervene++: An Adaptive Intervention Agent for Efficient Real-World Reinforcement Learning cs.LG · Oct 2026
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment