Robotics · arXiv:2610.28457 v1 · submitted 20 Oct 2026 · 12 references
GPT-6-Astra Understands Touch and Outperform Vision–Tactile–Language–Action Models
The paper says
“General-purpose multimodal models have shown promise as direct robot action policies, yet whether they grasp the physics underlying real-world interactions and can apply that understanding to guide action remains underexplored.”
more of the abstract
We investigate this question using contact-rich manipulation as a representative testbed, examining tactile understanding as a concrete instance of this broader capability. Contact-rich manipulation requires interpreting tactile feedback and adjusting actions after obstruction. Existing learned tactile policies acquire these capabilities through specialized pretraining and per-task optimization, with substantial data and training costs. We present GPT-as-Tactile-Policy, which summarizes expert demonstrations into task guidance and tactile references for contact assessment and corrective action. GPT-6-Astra combines this context with multimodal observations and interaction history to generate control commands and adjust actions through contact feedback. Without additional robotics-specific pretraining or per-task policy fine-tuning, our method achieves 93.8% mean success across five UniVTAC tasks. Overall performance is comparable to state-of-the-art policies, with higher success on selected tasks. On Insert Hole, tactile feedback improves success from 31% to 99%. Execution analysis further demonstrates contact-guided probing and correction, supporting manipulation progress through explicit expert experience and online feedback.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 21 Oct 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. A training-free tactile policy on UniVTAC that earns a serious referee, provided the authors close the protocol-equivalence and data-disjointness gaps before publication. This line is wrong
3 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 12 references say.
The paper says, at §4.1, Table 2
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The headline comparison to N0-VTLA (96.2%), N0-TWAM (84.6%), ACT + UniVTAC Encoder, and π0.5 rests on numbers quoted as 'previously evaluated on UniVTAC,' but the manuscript does not state that these baselines were evaluated on the same 100 fixed test episodes per task, the same simulator version, the same initial-state distribution, and the same success criteria as the proposed method.
more of objection 1
If the baseline figures come from a different evaluation setting, the central claim of comparable performance is not established. The authors should either cite the exact UniVTAC benchmark protocol that fixes these conditions for all methods, or rerun/report the baselines under the identical protocol.
The paper says, at §4.1, §3.2
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The Task Skill and tactile reference set are constructed from 40 expert demonstrations (with 10 used for offline validation), while evaluation uses 100 fixed test episodes per task. The paper does not state that the 100 test episodes are disjoint from the demonstrations used to build the task context, nor whether the UniVTAC benchmark guarantees this.
more of objection 2
Because the context directly injects content extracted from those demonstrations into the policy input, any overlap between test episodes and context-building demonstrations would inflate the reported success rates and would need to be ruled out explicitly.
The paper says, at §4.3, Table 3
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The tactile ablation compares 31% versus 99% success, but both numbers are point estimates from a single run of 100 configurations with no error bars, confidence intervals, or repeated-seed information; the largest reported effect is also at the ceiling. A binomial 95% confidence interval for 31/100 is roughly 22–41% and for 99/100 roughly 95–100%, so even the unquantified gap is …
more of objection 3
… likely large, but the claim 'tactile feedback improves success from 31% to 99%' needs an uncertainty estimate or repeated evaluation to be load-bearing. Reporting per-seed or per-configuration variation would also clarify whether the 68-point gap is stable.
4 minor objections
The paper says, at Title
Pith says: minor objection 1 of 4
The title contains a grammatical error: 'Outperform' should be 'Outperforms' or 'Outperforming.'
The paper says, at §2.1
Pith says: minor objection 2 of 4
In the sentence ending 'Following this direct action-decision paradigm, We study this direct-policy paradigm...', 'We' should be lowercase and the phrasing is redundant.
show 2 more
The paper says, at §1, §4.1
Pith says: minor objection 3 of 4
The contributions state evaluation 'over 100 seeds per task following the benchmark protocol,' while §4.1 states '100 fixed test episodes per task.' These should be reconciled, since 'fixed episodes' and 'seeds' describe different randomization procedures.
The paper says, at §4.3
Pith says: minor objection 4 of 4
For the w/o tactile variant, the text says it retains vision, proprioception, and the action interface while removing tactile information and tactile references, but Table 3 labels only the observation components; it would be clearer to state explicitly in the table caption that the variant also excludes the tactile references from the task context.
What would settle it
Request the exact task/seed/success-criterion specification used for each quoted baseline in Table 2, and check three things: (1) Are the 100 test episodes per task identical to the episode set on which N0-VTLA/N0-TWAM reported their numbers, or re-drawn from the same distribution? (2) Are those 100 episodes disjoint from the 40 knowledge-construction and 10 offline-validation demonstrations, and is the Task Skill frozen before the fixed test set is defined? (3) Re-run the Insert Hole ablation over at least 10 independently seeded episode sets of 100 episodes and report the mean and standard error of (Ours, w/o tactile); if the 99% vs 31% gap persists across seeds, the tactile claim holds. If the authors cannot supply matching protocol details, the claim should be downgraded to conditional on protocol equivalence.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 21 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 4 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
This is the rare GPT-as-policy paper worth taking seriously. The new idea is not the model—it is the packaging: a fixed general-purpose multimodal model, a textual Task Skill distilled from 40 expert demonstrations, a small tactile reference set, and closed-loop action generation. No representation pretraining, no per-task policy fine-tuning. If the numbers hold, this is a genuinely different route to task adaptation and makes tactile policy acquisition much cheaper than ACT+UniVTAC, N0-VTLA, or N0-TWAM. The claim is internally coherent: parameters are fixed, the benchmark is external, and there are no fitted parameters, so the circularity burden is low. The tactile ablation (31% to 99% on Insert Hole) is the strongest result in the paper.
The soft spots are where I would push back. The comparison is load-bearing and rests on quoted baseline numbers. The paper says it uses 100 fixed test episodes per task but does not demonstrate that N0-VTLA and N0-TWAM were evaluated on the same episodes, same simulator, or same success criteria. It also never states explicitly that the 100 test episodes are disjoint from the 40 demonstrations used to build the Task Skill. I do not think this is evidence of leakage—the Task Skill is summarized guidance, not episode memory—but the authors should state it plainly. The 31% vs 99% ablation is a single point estimate with no error bars or repeated runs; the gap is large enough to survive a little noise, but as reported it is less solid than it looks. No prompts, no code, and no examples of the actual Task Skill are released, so reproduction is currently not possible. A minor point: the same model distills the Task Skill and then follows it, so part of the measured success reflects instruction following; that is interesting rather than damning, but cross-model distillation would strengthen the claim.
One mismatch: the title says "Understands Touch and Outperform," while the abstract says comparable, and the average (93.8) is below N0-VTLA (96.2). Not a contradiction, but the title oversells.
Bottom line: I would send this to peer review. Serious referee time is justified. The main revisions I would ask for are protocol equivalence, episode disjointness, error bars, and release of the prompt and context materials. The paper is for people working on tactile VLAs, in-context policy learning, and benchmark protocol design. It does not close the loop on real-world contact-rich manipulation—simulation only—but it frames a testable question that the field should engage with.
The strongest objection
The central claim is that a fixed-parameter general-purpose model, given in-context expert summaries and tactile feedback, is comparable to or better than trained visuo-tactile policies on UniVTAC. The load-bearing comparison is Section 4.1 / Table 2: 'Our evaluation comprises 100 fixed test episodes per task' against ACT + UniVTAC Encoder, π0.5, N0-VTLA, and N0-TWAM 'previously evaluated on UniVTAC.' The paper does not state that the same 100 episodes, same simulator version, same initial-state distribution, and same success criteria were used for the quoted baselines, nor that the 100 test episodes are disjoint from the 50 expert demonstrations used to build the Task Skill (Sec. 4.1: 40 for knowledge construction, 10 for offline validation). Since the Task Skill is distilled from those demonstrations, any overlap with the fixed test set would mean the 'policy' has effectively seen test content in context; even without literal overlap, the authors do not report whether baseline protocols use the identical fixed episode set or fresh random draws. The second load-bearing result, the tactile ablation (Table 3), is a single 31% vs 99% comparison with no error bars, statistical test, or repeated-seed information; Section 4.3 says 'same 100 test configurations' once, which does not tell how robust the 68-point gap is. These are protocol-equivalence and sampling concerns, not internal contradictions; the training-free claim itself is well supported by the fixed-parameter statement (b85) and the method description, and the paper does not claim data release, so lack of release is not an attack on correctness.
The reader's weakest assumption (episode/protocol equivalence between the quoted baselines and the 100 fixed test episodes, plus possible leakage from demonstrations into the test set) is exactly the most load-bearing point. The paper's own protocol sentences (Section 4.1) explicitly state '100 fixed test episodes per task' and that baselines were 'previously evaluated on UniVTAC', but nowhere state that the episodes, simulator version, or success criteria coincide, nor that the fixed test episodes are disjoint from the demonstrations used to build the task context. Since the central comparison (93.8% vs 96.2% vs 84.6%) depends on cross-paper comparability, this is a genuine correctness risk rather than a scope disagreement. The tactile ablation strengthens the paper's qualitative claim, but its single evaluation of 100 episodes with no error bars leaves the magnitude (31% to 99%) as a point estimate. I do not see an internal contradiction that would justify REJECT, and the training-free construction is concretely described with fixed parameters (Equation 1, b85), so the paper is at least CONDITIONAL. The published-source rule applies to baselines quoted from cited works, so this is not an attack on those papers; it is a request to verify that the current paper applied the same protocol, which is a standard reporting obligation. The test is concrete, fair, and would settle the concern: protocol equivalence check plus a seeded replication of the ablation.
The referee's report
If the central claim holds, the paper makes a substantial point: a fixed-parameter general-purpose multimodal model, supplied with expert experience in context and online tactile feedback, can reach the level of trained visuo-tactile policies on a standard benchmark. The paper has real methodological strengths: evaluation is on an external benchmark (UniVTAC), the policy parameters and reasoning configuration are fixed across tasks, task context is fixed before formal evaluation, and the claim of no additional pretraining or per-task parameter updates is clearly stated. The tactile ablation is large in magnitude. The main risk is not internal circularity of the method itself, but whether the comparison to quoted baselines and the reported ablation are supported by sufficiently specified evaluation protocols.
Pith's reading of the paper (a machine's summary, not the authors')
This paper tries to establish that a general-purpose multimodal model can act directly as a robot's hands on contact-rich manipulation, understanding touch well enough to correct its own actions, with no robotics-specific pretraining and no per-task fine-tuning. The proposed method, GPT-as-Tactile-Policy, keeps GPT-6-Astra's parameters fixed and instead supplies expert experience in context: a textual Task Skill and reference tactile observations distilled from expert demonstrations, combined at every decision step with live bilateral fingertip images, indentation and deformation statistics, and the episode's action-feedback history. On five UniVTAC tasks, with 100 fixed test episodes per task, the paper reports 93.8% mean success, close to the 96.2% of the strongest trained baseline and ahead of an extensively pretrained world-action model, with tactile feedback raising Insert Hole success from 31% to 99%. If the claim holds, expert experience can be converted into a reusable prompt asset instead of learned weights, which would substantially cut the data and compute cost of tactile manipulation skills.
What carries the argument. The load-bearing object is the fixed task context $\mathcal{C}_\tau = (K_\tau, E^{\mathrm{tac}}_\tau)$, assembled offline by the same frozen model from expert demonstrations: $K_\tau$ is the textual Task Skill, which organizes the demonstrations into the task objective, manipulation principles, evidence for stage transitions, and corrective strategies after obstruction, while $E^{\mathrm{tac}}_\tau$ is a set of reference tactile observations preserving bilateral contact distributions and local deformation patterns. This context is concatenated with the current observation $o_t$ (scene and wrist RGB, proprioception, tactile images, indentation statistics, marker displacements) and the interaction history $\mathcal{H}_t$, and the policy draws each action as $a_t \sim \pi_\theta(\cdot \mid \ell, \mathcal{C}_\tau, o_t, \mathcal{H}_t)$ with parameters $\theta$ fixed. The mechanism replaces specialized tactile encoders and per-task policy training with two things a frozen model can consume directly: explicit prior experience stating what to check and how to react, and online contact feedback showing what the current attempt actually produced. Short execution horizons of one to five control steps make the feedback loop dense enough that each probing attempt can be evaluated and then abandoned, adjusted, or continued.
Core claim. The central claim is that one fixed-parameter general-purpose multimodal model, GPT-6-Astra, can serve as a complete closed-loop policy for contact-rich manipulation: at each step it receives the task instruction, a fixed task context built offline from expert demonstrations, scene and wrist camera views, proprioception, bilateral tactile images with indentation statistics and marker-displacement descriptors, and the episode's interaction history, and it emits a numerical end-effector pose, gripper command, and a short execution horizon that a generic inverse-kinematics controller carries out. All policy parameters stay frozen; the only task-specific asset is the in-context task context. On the five UniVTAC tasks — Insert Hole, Lift Can, Grasp Classify, Lift Bottle, and Pull-out Key — the paper reports a mean success rate of 93.8% over 100 fixed test episodes per task, approaching the 96.2% of the strongest trained baseline, exceeding the extensively pretrained world-action baseline on every task, and beating the ACT + UniVTAC encoder and the vision-only baseline outright. On Insert Hole the method reaches 99%, matching the best trained policy, and removing tactile input drops it to 31%, a 68-point gap that the paper reads as evidence that touch, not vision, carries contact-guided correction. On Lift Can it reaches 95%, the highest rate among the compared methods.
The premise it rests on. The load-bearing premise is that the 100 fixed test episodes per task are an independent, fair sample — in particular disjoint from the 50 expert demonstrations used to build the task context — and that the quoted baseline success rates were produced under the same episode set, simulator version, and success criteria; the paper states its own protocol but does not explicitly establish either condition.
If the paper is right (Pith's inferences, not the authors')
If the 93.8% figure holds, adapting to a new contact-rich task requires only the offline knowledge-construction pass over expert demonstrations, not gradient updates, so the per-task cost of tactile skill collapses to data collection and prompt building.
The 31% to 99% swing on Insert Hole implies that for tasks with hidden geometric constraints, tactile feedback is the decisive input even for a model with strong visual priors, an argument for equipping generalist robot policies with contact sensors rather than relying on vision alone.
Because the Task Skill is human-readable text, the grounds for a probing or correction decision are in principle inspectable and editable, a form of policy transparency that trained tactile networks do not offer.
Since one fixed model with one reasoning configuration handles all five tasks, the result points toward a single frozen backbone plus per-task context as a common policy interface for a benchmark that previously required one trained policy per task.
Editorial inference: the recipe likely generalizes past touch — any feedback channel whose expert demonstrations can be summarized into reference snapshots plus decision rules (force-torque, proximity, acoustic, thermal) could be plugged into the same frozen-model loop, turning sensor choice into a prompt-design problem.
Editorial inference: the paper does not separate two possible sources of the gain — genuine online interpretation of contact physics versus retrieval of corrective patterns memorized from the demonstrations used to build the Task Skill; testing on object geometries and obstruction directions absent from the demonstrations would split the two.
Editorial inference: if the method transfers to real hardware, the practical bottleneck shifts from compute to data curation, because the quality and coverage of the expert demonstrations, and the faithfulness of the prompt to them, would likely set the ceiling on new tasks.
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
General-purpose multimodal models have shown promise as direct robot action policies, yet whether they grasp the physics underlying real-world interactions and can apply that understanding to guide action remains underexplored. We investigate this question using contact-rich manipulation as a representative testbed, examining tactile understanding as a concrete instance of this broader capability. Contact-rich manipulation requires interpreting tactile feedback and adjusting actions after obstruction. Existing learned tactile policies acquire these capabilities through specialized pretraining and per-task optimization, with substantial data and training costs. We present GPT-as-Tactile-Policy, which summarizes expert demonstrations into task guidance and tactile references for contact assessment and corrective action. GPT-6-Astra combines this context with multimodal observations and interaction history to generate control commands and adjust actions through contact feedback. Without additional robotics-specific pretraining or per-task policy fine-tuning, our method achieves 93.8% mean success across five UniVTAC tasks. Overall performance is comparable to state-of-the-art policies, with higher success on selected tasks. On Insert Hole, tactile feedback improves success from 31% to 99%. Execution analysis further demonstrates contact-guided probing and correction, supporting manipulation progress through explicit expert experience and online feedback.
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 · 2026-10-21
The authors' main claim, as the classifier picked it out 0.52
“Without additional robotics-specific pretraining or per-task policy fine-tuning, our method achieves 93.8\% mean success across five UniVTAC tasks.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.20 · human participants 0.09 · controlled comparison 0.60 · randomized assignment 0.01 · states quantitative result 0.99 · compares with prior results 0.96 · mentions code or data availability 0.03
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 |
|---|---|---|---|---|
| General-purpose multimodal models have shown promise as direct robot action policies, yet whether they grasp the physics underlying real-world interactions and can apply that understanding to guide action remains underexplored. | gap 0.87 | 0.04 | 0.04 | asserted 0.70 |
| We investigate this question using contact-rich manipulation as a representative testbed, examining tactile understanding as a concrete instance of this broader capability. | contribution 0.57 | 0.68 | 0.03 | asserted 0.74 |
| Contact-rich manipulation requires interpreting tactile feedback and adjusting actions after obstruction. | background 1.00 | 0.06 | 0.03 | asserted 1.00 |
| Existing learned tactile policies acquire these capabilities through specialized pretraining and per-task optimization, with substantial data and training costs. | background 0.96 | 0.03 | 0.03 | asserted 1.00 |
| We present GPT-as-Tactile-Policy, which summarizes expert demonstrations into task guidance and tactile references for contact assessment and corrective action. | method 0.57 | 0.90 | 0.07 | asserted 1.00 |
| GPT-6-Astra combines this context with multimodal observations and interaction history to generate control commands and adjust actions through contact feedback. | method 0.99 | 0.75 | 0.08 | asserted 1.00 |
| Without additional robotics-specific pretraining or per-task policy fine-tuning, our method achieves 93.8\% mean success across five UniVTAC tasks. | result 1.00 | 0.96 | 0.37 | asserted 1.00 |
| Overall performance is comparable to state-of-the-art policies, with higher success on selected tasks. | result 0.95 | 0.89 | 0.26 | asserted 1.00 |
| On Insert Hole, tactile feedback improves success from 31\% to 99\%. | result 1.00 | 0.92 | 0.13 | asserted 1.00 |
| Execution analysis further demonstrates contact-guided probing and correction, supporting manipulation progress through explicit expert experience and online feedback. | interpretation 0.75 | 0.91 | 0.10 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 0.99 · evidence from observation or experiment 0.99
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, 2026-10-21: 1 request, 7,440 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
Figures (4)
Figures are the authors', reproduced from the paper.
References (12)
- [4] B. Huang, Y. Wang, X. Yang, Y. Luo, and Y. Li. 3d-vitac: Learning fine-grained manipulation with visuo-tactile sensing.arXiv preprint arXiv:2410.24091, on PithPith's reviewarXiv
- [6] S. Huang, Z. Jiang, H. Dong, Y. Qiao, P . Gao, and H. Li. Instruct2act: Mapping multi-modality instructions to robotic actions with large language model.arXiv preprint arXiv:2305.11176, 2023a. W. Huang, F. Xia, T. Xiao, H. Chan, J. Liang, P . Florence, A. Zeng, J. Tompson, I. ... on PithPith's reviewarXiv
- [8] Z. Liu, A. Bahety, and S. Song. Reflect: Summarizing robot experiences for failure explanation and correction.arXiv preprint arXiv:2306.15724, on PithPith's reviewarXiv
- [9] Y. Mikami, A. Melnik, J. Miura, and V . Hautamäki. Natural language as policies: Reasoning for coordinate-level embodied control with llms.arXiv preprint arXiv:2403.13801, on PithPith's reviewarXiv
- [10] G. Team, X. Chen, X. Cheng, Y. Deng, L. Ding, S. Dong, X. Gao, H. Jia, Z. Li, Z. Li, et al. Systematically exploring the capabilities of gpt-6 astra as embodied policies.arXiv preprint arXiv:2609.38537, 2026a. N. Team, F. T. Team, et al.𝑛_0-twam: Scaling tactile-native world-a... on PithPith's reviewarXiv
- [11] S. Wu, L. You, J. Zhu, Y. Liu, H. Kaixiang, C. Yonghang, J. Li, C. Zhang, J. Liu, H. Chu, et al. Tactile-wam: Touch-aware world action model with tactile asymmetric attention.arXiv preprint arXiv:2606.26663, on PithPith's reviewarXiv
- [12] H. Xue, J. Ren, W. Chen, G. Zhang, Y. Fang, G. Gu, H. Xu, and C. Lu. Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation.arXiv preprint arXiv:2503.02881, on PithPith's reviewarXiv
- [2022] 12 W. Huang, C. Wang, R. Zhang, Y. Li, J. Wu, and L. Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models.arXiv preprint arXiv:2307.05973, 2023b. P . Intelligence, K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. Equi... on PithPith's reviewarXiv
- [2023] B. Chen, W. Wan, T. Chen, X. Guo, C. Xu, Y. Qi, H. Zhang, L. Wu, T. Xu, Z. Li, et al. Univtac: A unified simulation platform for visuo-tactile manipulation data generation, learning, and benchmarking.arXiv preprint arXiv:2602.10093, on PitharXiv
- [2024] J. Huang, S. Wang, F. Lin, Y. Hu, C. Wen, and Y. Gao. Tactile-vla: unlocking vision-language- action model’s physical knowledge for tactile generalization.arXiv preprint arXiv:2507.09160, on PithPith's reviewarXiv
- [2025] C. Higuera, S. Arnaud, B. Boots, M. Mukadam, F. R. Hogan, and F. Meier. Visuo-tactile world models.arXiv preprint arXiv:2602.06001, on PitharXiv
- [2026] J. Chen, Y. Mu, Q. Yu, T. Wei, S. Wu, Z. Yuan, Z. Liang, C. Yang, K. Zhang, W. Shao, et al. Roboscript: Code generation for free-form manipulation tasks across real and simulation. arXiv preprint arXiv:2402.14623, on PithPith's reviewarXiv
Cited by 2 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models cs.RO · Nov 2026
- Embodied First Exam cs.RO · Nov 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