Pith.

Robotics · arXiv:2611.08156 v1 · submitted 5 Nov 2026 · 57 references

La-JEPA: Compact Robot Policies with Latent World Modeling

Mateo Clémente, George Z. Xue, Amin Abyaneh, Liheng Ma, Rui Heng Yang, Tongtong Cao, Zhanguang Zhang, Yingxue Zhang

The paper says

“World-action models adapt pretrained video generators, and vision-language-action models build on pretrained vision-language backbones; both tie policy architecture and inference cost to large transformers designed for generation rather than control. It remains unclear how much these generation-oriented choices affect control.”

more of the abstract

In this work, we introduce Latent Action JEPA (La-JEPA), a compact world-action model whose policy transformer is trained from scratch in the feature space of a pretrained V-JEPA 2.1 encoder. Following Fast-WAM, La-JEPA learns to predict future features jointly with actions during training and generates only actions at inference. Among policies without embodied pretraining, La-JEPA achieves state-of-the-art performance in clean and randomized RoboTwin 2.0 scenes and on LIBERO-Plus with correct instructions, while using far fewer parameters than the strongest competing policies. It runs substantially faster than Fast-WAM at comparable performance and outperforms the closest JEPA-based methods, LeapBot-WA and JEPA-WAM, on LIBERO-Plus and RoboTwin 2.0 without their pretrained policy backbones. On four real-robot tasks, its average success is comparable to pi0.5 and above non-embodied-pretrained baselines. Through a design study of the choices behind La-JEPA, we show that JEPA features outperform video VAE latents, most clearly out of distribution, and that training-time future prediction choices significantly impact the policy even when future prediction is skipped at inference. We correct the LIBERO-Plus instructions, as the officially released evaluator corrupts them for most variants, and re-evaluate several strong open baselines. Without a generative pretrained policy backbone, La-JEPA is better and faster than far larger policies, suggesting that robot foundation models may benefit from architectures designed for control rather than inherited from generation.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 6 Nov 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. Solid, transparent systems paper: the RoboTwin claims are robust, but the LIBERO-Plus SOTA hinges on the corrected-instruction protocol, which is plausible yet not externally confirmed; worth a serious referee. This line is wrong

2 major objections · 4 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 57 references say.

The paper says, at §5.2, Table 14, Abstract (claims b4 and b20)

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.67):

“Because the corruption affects policies to different degrees, released-instruction scores partly rank policies by their sensitivity to corrupted text rather than by their robustness to the intended perturbation: correcting the instructions reverses the order of NowW AM and La-JEPA B and L, and of Cosmos Policy and La-JEPA S.”

Pith says: Objection 1 of 2 · major

The LIBERO-Plus state-of-the-art claim in the abstract and introduction is conditioned on the corrected-instruction protocol, but that protocol is an interpretive claim about benchmark intent rather than an established fact. Under the officially released evaluator, La-JEPA L is second to NowWAM (84.1 vs 86.2), and the order of Cosmos Policy and La-JEPA S inverts (Table 14; the paper itself notes the inversion at sentence b360).

more of objection 1

The correction is plausible and transparently documented (the released evaluator appends perturbation parameters to 8,493 of 10,030 instructions), but Flex-π's different choice of BDDL paraphrases, described in Section 5.2, shows that the community has not converged on a single protocol, and the LIBERO-Plus authors' intent is not established by any external source. Because the headline claim rests on this premise, the manuscript should either (a) re-scope the headline so that the protocol sensitivity is a central caveat rather than a side note, or (b) add a protocol-robustness analysis, for example evaluating La-JEPA and the re-evaluated baselines under Flex-π's annotation protocol as well, or (c) obtain confirmation from the LIBERO-Plus authors that the verbatim base-task instruction is the intended text. I agree with the paper's observation in Section 5.2 that released-instruction scores partly rank policies by their sensitivity to corrupted text, but that is an argument, not an established fact, and it is load-bearing for the central claim.

The paper says, at Table 1 note, §5.2, Table 14

The review did not quote this passage.

Pith says: Objection 2 of 2 · major

The claimed lead over NowWAM is sensitive to the evaluation harness in two ways that the framing does not fully convey. First, the corrected-instruction comparison (88.0 vs 86.2) is measured against a NowWAM re-scored at upstream per-suite step limits, whereas NowWAM's published 87.7% uses its own longer limits (400 steps, 700 for LIBERO-Long), as the Table 1 note discloses; under NowWAM's native step protocol the +1.8-point lead could shrink or invert.

more of objection 2

Second, the instruction correction itself moves La-JEPA L by +3.9 points (84.1 to 88.0) while NowWAM is unchanged (86.2 to 86.2), so the margin is of the same magnitude as the protocol-dependent swing. The manuscript should report the corrected-instruction comparison under both step-limit settings, and it should temper the Section 5.2 statement that 'L leads NowWAM by 1.8 points with four times fewer deployed parameters' by explicitly bounding the claim with the harness used.

4 minor objections

The paper says, at Appendix B (JEPA-WAM full-suite estimate), Figure 4, §5.3

Pith says: minor objection 1 of 4

The JEPA-WAM* markers in Figure 4 and the supporting statement in Section 5.3 come from linear regressions fit on seven policies that include La-JEPA S and B; although the appendix states this plainly, the main text and figure caption should repeat the caveat, because the asterisked point is otherwise easily read as a measured score.

The paper says, at Section 5.4 and Table 15

Pith says: minor objection 2 of 4

The latency comparison mixes each method's native configuration: La-JEPA uses two 256×256 two-frame cameras with FLAN-T5-large timed in FP32, while Fast-WAM and Faster-WAM use two 224×224 single-frame cameras with UMT5-XXL in BF16. The protocol is documented in Appendix C.1, but the 'up to 5× faster' phrasing in Section 5.4 should point the reader to this cross-configuration caveat at the point of the claim.

show 2 more

The paper says, at Appendix B (Data and evaluation pipelines)

Pith says: minor objection 3 of 4

The revised LIBERO pipeline trains on 1,777 of the 2,000 demonstrations and the RoboTwin pipeline on 2,438 of 2,500, whereas the compared baselines trained on the full released sets; since Table 1 is a head-to-head state-of-the-art comparison, a sentence should state whether the filtering is expected to have any systematic effect on training difficulty and hence on the comparison.

The paper says, at Table 2

Pith says: minor objection 4 of 4

The Faster-WAM RoboTwin row is the authors' retraining on clean demonstrations rather than a score reported in the Faster-WAM paper; the table would benefit from a footnote marking the entry as retrained by the authors, since that information currently appears only in Appendix B.

What would settle it

Obtain a ruling from the LIBERO-Plus authors (or an official corrected-evaluator release) confirming that non-language variant instructions should be the base task instruction verbatim, with the perturbation expressed only through scene/state changes; then have a third party re-run the Table 14 checkpoints of NowWAM and La-JEPA L under that canonical protocol with identical harness settings. If the corrected protocol is confirmed and the reported margin persists, the SOTA claim holds; if the released evaluator is judged canonical, the claim reduces to 'second to NowWAM'.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 6 Nov 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 36 s · 33,231 tokens read, 4,777 written
  3. Editor wrote the verdict above 20 s · 34,879 tokens read, 772 written
  4. Referee major revision: 2 major objections, 4 minor 1 min 36 s · 33,644 tokens read, 15,135 written

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

Read this one. It is a genuinely well-done empirical systems paper, and the transparency is unusual: three training seeds with standard deviations, both released and corrected instruction scores reported, per-task tables, and a limitations section that names most of its own weaknesses before you can. The central idea — a compact world-action policy trained from scratch in a frozen V-JEPA 2.1 feature space, no generative backbone — is new as a combination, and the design study produces reusable findings. The JEPA-over-VAE result (most clearly out of distribution) and the result that future prediction during training helps even when skipped at inference are the kind of measurements people will actually reuse.

The RoboTwin 2.0 claim is the strong one: all three sizes beat every non-embodied baseline in both clean and randomized scenes, with the margin largest exactly where you want generalization evidence. That part does not depend on any protocol dispute.

The soft spot is the LIBERO-Plus SOTA, and the stress-test note lands on it correctly. Under the released instructions, La-JEPA L is second to NowWAM (84.1 vs 86.2), and the ordering of Cosmos Policy and La-JEPA S inverts too. The paper's SOTA claim holds only under its corrected-instruction protocol, which replaces the released evaluator's corrupted texts with the base training instructions verbatim. The corruption is real and documented — 8,493 of 10,030 variants get perturbation parameters appended — and the correction is disclosed thoroughly. But it is an interpretive claim about benchmark intent, not something LIBERO-Plus's authors have blessed, and Flex-pi's different response shows the community has not converged. Also worth noting: NowWAM is scored at upstream step limits rather than its own longer limits, which it reports as 87.7. These are disclosed choices, not tricks, but they make the SOTA claim condition-dependent. The margin over NowWAM is 1.8 points with corrected instructions, and it would not take much to flip.

Secondary weaknesses are smaller. Some design-study blocks compare recipes rather than isolated factors — the authors admit this, in the text and in the appendices. The JEPA-WAM 20-to-50 task extrapolation is a fitted relation, but it is excluded from result tables and not load-bearing. Real-robot evidence is 20 rollouts per task on one setup, acknowledged.

As a referee I would want an independent run of the corrected-instruction evaluations before shipping the headline, but this paper deserves that referee time. Send it out.

The strongest objection

The central claim (b8, restated in b4 and b20) is that La-JEPA achieves state-of-the-art LIBERO-Plus performance among policies without embodied pretraining. In the authors' own Table 14, this holds only under their corrected-instruction protocol: with the officially released instructions, La-JEPA L scores 84.1 against NowWAM's 86.2, and the order of Cosmos Policy and La-JEPA S also inverts (b360). The correction is plausible and transparently documented — the released evaluator appends perturbation parameters to 8,493 of 10,030 instructions — but the paper's justification that this 'corrupts' the benchmark and that the verbatim base-task instruction is the intended protocol is an interpretive claim about benchmark intent, not a fact established by the LIBERO-Plus authors. Flex-π's different response to the same issue (BDDL paraphrases) shows that the community has not converged on one protocol. Additionally, the head-to-head margin over NowWAM is sensitive to the harness: NowWAM's published 87.7 on released instructions uses longer step limits, while the table scores it at upstream limits (Table 1 note). None of this suggests unfairness — the paper reports both protocols expressly — but the SOTA claim's truth is benchmark-definition-dependent, and the condition that would make it true remains externally unconfirmed.

The reader's weakest_assumption identifies exactly the premise I find most load-bearing: the LIBERO-Plus SOTA statement is conditional on the corrected-instruction protocol being the intended benchmark. This is not a manufactured concern; the paper's own Table 14 shows the ranking inverts under the released evaluator, so a well-specified, externally confirmable premise determines whether the headline claim is true. The authors deserve credit for disclosing both sets of numbers, reporting the correction in detail, and re-evaluating the open baselines under one shared harness; the remaining risk is not about honesty or internal consistency but about benchmark definition. I considered other candidates — the 20-rollout real-robot comparison, the non-isolating design-study comparisons (b253, b286), and the non-monotonic RoboTwin scaling (b119, b147) — but each is either explicitly acknowledged as a limitation or does not carry the central claim the way the LIBERO-Plus protocol does. The RoboTwin result (Table 2) is protocol-independent in the same sense and is the strongest independent support for the paper's thesis, but the abstract's LIBERO-Plus SOTA statement still leans on the correction. Since the reader already judged the paper CONDITIONAL and my analysis converges on the same premise, no verdict adjustment is needed.

The referee's report

If the claims hold, this is a significant result for the design of robot foundation models: it demonstrates that a compact, from-scratch policy transformer operating in a pretrained JEPA feature space can match or beat generative-backbone policies several times its size, and the design study gives the community reusable evidence about where the gains come from. The experimental protocol is unusually transparent: headline numbers are three-seed means with standard deviations; both released and corrected LIBERO-Plus scores are reported; real-robot results use 20 rollouts per task; open baselines are re-evaluated in a single harness with documented step limits; and the code is released. The authors also state their limitations honestly, including non-monotonic RoboTwin scaling, confounds in some ablation comparisons, and the narrow single-robot scope of the real-robot study.

Pith's reading of the paper (a machine's summary, not the authors')

La-JEPA asks whether a robot policy must inherit the architecture of the pretrained video or language generators that today's world-action and vision-language-action models are built on. The paper answers no: a small diffusion transformer trained from scratch, operating in the frozen feature space of a V-JEPA 2.1 video encoder, reports state-of-the-art success among policies without embodied pretraining on RoboTwin 2.0 and on LIBERO-Plus, using 0.74 to 1.82 billion deployed parameters against 7 to 12 billion for the strongest baselines and running up to five times faster than the closest design it follows. The LIBERO-Plus claim comes with a disclosed catch: the official evaluator appends perturbation parameters to most instructions, so the paper corrected the instruction texts to isolate each perturbation as intended, and under the released instructions La-JEPA L is second to NowWAM rather than first. If the claim holds, robot foundation models may be better built for control than inherited from generation, with world modeling kept as a training-time signal that costs nothing at inference.

What carries the argument. The load-bearing object is the pretrained V-JEPA 2.1 video encoder's dense feature space, into which observations and future frames are encoded once, plus a pair of diffusion transformers trained from scratch: a World DiT that predicts future observation features and an Action DiT that generates the action chunk, both supervised by flow matching with a training objective summing two mean-squared-error terms, one on the predicted future-feature velocity and one on the predicted action velocity, under an attention mask that keeps future tokens out of the action pathway. At inference the future tokens are dropped, the World DiT's observation keys and values are cached once, and only the Action DiT denoises, which is what makes the policy fast. The design study isolates the mechanisms behind the recipe: JEPA features over video VAE latents (clearest out of distribution), tube-motion temporal encoding of two-frame clips, camera and proprioception dropout, and image augmentation with LoRA adaptation of the observation encoder.

Core claim. The paper's central claim is that a compact transformer trained from scratch in a good latent space can match or beat far larger policies whose predictors are inherited from video or language generation, and that the latent space and the predictive supervision, not the generative backbone, do the work. Concretely, the authors report that all three La-JEPA sizes outperform every compared policy without embodied pretraining on clean and randomized RoboTwin 2.0 scenes, that on LIBERO-Plus with corrected instructions La-JEPA L leads the strongest re-evaluated baselines while deploying about four times fewer parameters, that La-JEPA runs up to five times faster than Fast-WAM at comparable performance, and that its mean success across four real-robot Franka tasks is comparable to the embodied-pretrained pi0.5. In the design study, substituting JEPA features for Wan2.2 VAE latents improves success in every tested comparison and most clearly out of distribution, where the VAE-based policy collapses to about 1 percent randomized success on RoboTwin; joint future-feature prediction at training time improves success even though the future is never generated at inference; and camera dropout, proprioception dropout, augmentation, and encoder LoRA each contribute. The authors also show that the released LIBERO-Plus evaluator corrupts instructions for 8,493 of its 10,030 variants and that the corruption shifts reported scores by up to 19 points for some baselines, so they re-evaluate every open policy under corrected instructions in a single harness.

The premise it rests on. The LIBERO-Plus state-of-the-art claim rests on the assumption that the corrected-instruction protocol, which gives every non-language variant its base training instruction verbatim, is the intended benchmark and that all baselines are scored fairly in the authors' harness; under the officially released instructions, the paper reports La-JEPA L behind NowWAM (84.1% versus 86.2%), so that version of the claim would not hold.

If the paper is right (Pith's inferences, not the authors')

If the results hold, robot foundation models do not need generative backbones: a control-sized transformer in a good latent space reaches parity with four-to-ten-times-larger generation-based policies, which would cut the compute and memory cost of deployed manipulation policies.

World modeling can be a training-time-only signal: future-feature prediction improves success on LIBERO-Plus and RoboTwin yet is dropped at inference, giving the robustness of a world model at the latency of an action-only policy.

The latent space choice dominates out of distribution: JEPA features keep randomized-scene success high where VAE-latent policies collapse, so future world-action models should be built on task-oriented representations rather than pixel-reconstruction latents.

Published LIBERO-Plus rankings partly reflect sensitivity to corrupted instruction text rather than robustness to the intended perturbation, so the corrected re-scoring changes several ranking orders and should be the basis for future comparisons on that benchmark.

A direct test the paper leaves open: pretrain La-JEPA on large robot datasets and see whether randomized-scene success approaches the embodied-pretrained leaders; if it does, the gap to those methods was data, not capacity, which would change where the field spends its compute.

The training-time-only world model recipe is architecture-agnostic and could be grafted onto other closed-loop control problems with video observations, such as teleoperation or driving, to get predictive robustness without generative inference cost.

The instruction-corruption finding suggests auditing other perturbation benchmarks for the same failure mode: any evaluator that concatenates perturbation parameters into the prompt silently turns single-factor robustness tests into double perturbations and can misrank policies.

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

Every value above, in the review's public record (JSON)

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.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

World-action models adapt pretrained video generators, and vision-language-action models build on pretrained vision-language backbones; both tie policy architecture and inference cost to large transformers designed for generation rather than control. It remains unclear how much these generation-oriented choices affect control. In this work, we introduce Latent Action JEPA (La-JEPA), a compact world-action model whose policy transformer is trained from scratch in the feature space of a pretrained V-JEPA 2.1 encoder. Following Fast-WAM, La-JEPA learns to predict future features jointly with actions during training and generates only actions at inference. Among policies without embodied pretraining, La-JEPA achieves state-of-the-art performance in clean and randomized RoboTwin 2.0 scenes and on LIBERO-Plus with correct instructions, while using far fewer parameters than the strongest competing policies. It runs substantially faster than Fast-WAM at comparable performance and outperforms the closest JEPA-based methods, LeapBot-WA and JEPA-WAM, on LIBERO-Plus and RoboTwin 2.0 without their pretrained policy backbones. On four real-robot tasks, its average success is comparable to pi0.5 and above non-embodied-pretrained baselines. Through a design study of the choices behind La-JEPA, we show that JEPA features outperform video VAE latents, most clearly out of distribution, and that training-time future prediction choices significantly impact the policy even when future prediction is skipped at inference. We correct the LIBERO-Plus instructions, as the officially released evaluator corrupts them for most variants, and re-evaluate several strong open baselines. Without a generative pretrained policy backbone, La-JEPA is better and faster than far larger policies, suggesting that robot foundation models may benefit from architectures designed for control rather than inherited from generation.

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 · 2026-11-06

The authors' main claim, as the classifier picked it out 1.00

“Without a generative pretrained policy backbone, La-JEPA is better and faster than far larger policies, suggesting that robot foundation models may benefit from architectures designed for control rather than inherited from generation.”

How the main claim is stated: hedged possibility 0.87

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.26 · human participants 0.04 · controlled comparison 0.94 · randomized assignment 0.08 · states quantitative result 0.25 · 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.

SentenceRoleThe authors' own workStates the main conclusionStrength
World-action models adapt pretrained video generators, and vision-language-action models build on pretrained vision-language backbones; both tie policy architecture and inference cost to large transformers designed for generation rather than control. background 0.98 0.06 0.05 asserted 1.00
It remains unclear how much these generation-oriented choices affect control. gap 1.00 0.06 0.03 unclear 0.61
In this work, we introduce Latent Action JEPA (La-JEPA), a compact world-action model whose policy transformer is trained from scratch in the feature space of a pretrained V-JEPA 2.1 encoder. contribution 0.85 0.93 0.09 asserted 0.99
Following Fast-WAM, La-JEPA learns to predict future features jointly with actions during training and generates only actions at inference. method 0.99 0.82 0.05 asserted 1.00
Among policies without embodied pretraining, La-JEPA achieves state-of-the-art performance in clean and randomized RoboTwin 2.0 scenes and on LIBERO-Plus with correct instructions, while using far fewer parameters than the strongest competing policies. result 1.00 0.95 0.21 asserted 1.00
It runs substantially faster than Fast-WAM at comparable performance and outperforms the closest JEPA-based methods, LeapBot-WA and JEPA-WAM, on LIBERO-Plus and RoboTwin 2.0 without their pretrained policy backbones. result 1.00 0.94 0.15 asserted 1.00
On four real-robot tasks, its average success is comparable to pi0.5 and above non-embodied-pretrained baselines. result 1.00 0.94 0.10 asserted 1.00
Through a design study of the choices behind La-JEPA, we show that JEPA features outperform video VAE latents, most clearly out of distribution, and that training-time future prediction choices significantly impact the policy even when future prediction is skipped at inference. result 0.77 0.96 0.17 asserted 1.00
We correct the LIBERO-Plus instructions, as the officially released evaluator corrupts them for most variants, and re-evaluate several strong open baselines. contribution 0.50 0.95 0.03 asserted 0.97
Without a generative pretrained policy backbone, La-JEPA is better and faster than far larger policies, suggesting that robot foundation models may benefit from architectures designed for control rather than inherited from generation. interpretation 0.97 0.89 0.80 hedged possibility 0.87

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.96

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, 2026-11-06: 1 request, 7,794 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 (6)
Figure 1 of the paper
Figure 1. Success against deployed parameters, including frozen encoders, on LIBERO-Plus with
Figure 2 of the paper
Figure 2. La-JEPA. Left: general architecture. Right: context and future encoding with V-JEPA 2.1.
Figure 3 of the paper
Figure 3. Gain in success (percentage points) from each design choice over its baseline: A, action
Figure 4 of the paper
Figure 4. RoboTwin clean versus random￾ized success of clean-trained policies. Blue: La-JEPA B; brown: no embodied pretrain￾ing; gray: embodied pretraining; RoboManip: Qwen-RobotManip. Points above the dashed line through La-JEPA have a higher mean of the two regimes. ∗: JEPA-WAM estimate (Appendix B); LDA-1B as reported by Yan et al. (2026b). axes such as camera viewpoint, lighting, and sensor noise. On RoboTwin, all three La-JEPA sizes outperform every baseline in both regimes, with the largest margin in randomized scenes (
Figure 5 of the paper
Figure 5. Real-robot tasks: cube placement, fruit picking, Push T, and dish manipulation.
Figure 6 of the paper
Figure 6. Companion to Figure 1: in-distribution (top) and out-of-distribution (bottom) suc

Figures are the authors', reproduced from the paper.

References (57)
  1. [5] Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, Weiliang Deng, Yubin Guo, Tian Nian, Xuanbing Xie, Qiangyu Chen, Kailun Su, Tianling Xu, Guodong Liu, Mengkang Hu, Huan-ang Gao, Kaixuan Wang, Zhixuan L... on PithPith's reviewarXiv
  2. [9] Mateo Cl \'e mente, L \'e o Brunswic, Rui Heng Yang, Xuan Zhao, Yasser Khalil, Haoyu Lei, Amir Rasouli, and Yinchuan Li. Two-steps diffusion policy for robotic manipulation via genetic denoising. In Advances in Neural Information Processing Systems, 2025 on Pith
  3. [12] GigaBrain Team , Angen Ye, Axiang Sun, Can Jin, Chenxi Cheng, Chong Shi, Dengke Shang, Dingqian Zhang, Guan Huang, Guangqiang Wang, Guangqing Ding, Guo Li, Hangcong Li, Hengyu Zhong, Hongtao Lu, Jianbo Qin, Jiming Mao, Jing Zhu, Jindi Lv, Jingzhi Cui, Junjie Xie, Junyi Bao, Ka... on PithPith's reviewarXiv
  4. [17] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success, 2025 on Pith
  5. [18] Moo Jin Kim, Yihuai Gao, Tsung-Yi Lin, Yen-Chen Lin, Yunhao Ge, Grace Lam, Percy Liang, Shuran Song, Ming-Yu Liu, Chelsea Finn, and Jinwei Gu. Cosmos Policy: Fine-Tuning Video Models for Visuomotor Control and Planning , 2026. arXiv:2601.16163 on PithPith's reviewarXiv
  6. [25] Jiangran Lyu, Kai Liu, Xuheng Zhang, Haoran Liao, Yusen Feng, Wenxuan Zhu, Tingrui Shen, Jiayi Chen, Jiazhao Zhang, Yifei Dong, Wenbo Cui, Senmao Qi, Shuo Wang, Yixin Zheng, Mi Yan, Xuesong Shi, Haoran Li, Dongbin Zhao, Ming-Yu Liu, Zhizheng Zhang, Li Yi, Yizhou Wang, and He W... on PithPith's reviewarXiv
  7. [30] Physical Intelligence , Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin L... on PithPith's reviewarXiv
  8. [31] Qwen , An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le... on PithPith's reviewarXiv
  9. [33] StarVLA Community . StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing , 2026. arXiv:2604.05014 on PithPith's reviewarXiv
  10. [36] Team Wan , Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pand... on PithPith's reviewarXiv
  11. [42] Haoqi Yuan, Zhixuan Liang, Anzhe Chen, Ye Wang, Haoyang Li, Pei Lin, Yiyang Huang, Zixing Lei, Tong Zhang, Jiazhao Zhang, Jie Zhang, Jingyang Fan, Gengze Zhou, Qihang Peng, Chenxu Lv, Xiaoyue Chen, An Yang, Fei Huang, Junyang Lin, Dayiheng Liu, Jingren Zhou, Chenfei Wu, and Xi... on PithPith's reviewarXiv
  12. [47] Tianyuan Yuan and Zibin Dong and Yicheng Liu and Hang Zhao , year =. 2603.16666 , archivePrefix = on PithPith's reviewarXiv
  13. [48] Liheng Ma and Rui Heng Yang and Zhanguang Zhang and Mateo Clemente and Ziwen Hu and Tongtong Cao and Yingxue Zhang , year =. 2608.02365 , archivePrefix = on PithPith's reviewarXiv
  14. [49] Moo Jin Kim and Yihuai Gao and. 2026 , eprint = on Pith
  15. [50] Yuyang Zhang and Wenyao Zhang and Zekun Qi and He Zhang and Haitao Lin and Jingbo Zhang and Yao Mu and Xiaokang Yang and Wenjun Zeng and Xin Jin , year =. 2606.19531 , archivePrefix = on PithPith's reviewarXiv
  16. [51] Haoqi Yuan and Zhixuan Liang and Anzhe Chen and Ye Wang and Haoyang Li and Pei Lin and Yiyang Huang and Zixing Lei and Tong Zhang and Jiazhao Zhang and Jie Zhang and Jingyang Fan and Gengze Zhou and Qihang Peng and Chenxu Lv and Xiaoyue Chen and An Yang and Fei Huang and Junya... on Pith
  17. [52] Yihan Lin and Jiawei He and Shifeng Bao and Chen Zhao and Yang Li and Xiaobo Wang and Yan Wang and Cheng Chi and Jing Zhang , year =. 2608.09381 , archivePrefix = on PithPith's reviewarXiv
  18. [53] Pei Liu and Nan Zheng and Lang Zhang and Daojie Peng and Yanan Zhang and Feilong Kong and Mingyue Feng and Jiachao Liu and Yaonong Wang and Qifeng Chen and Jun Ma , year =. 2607.23969 , archivePrefix = on PithPith's reviewarXiv
  19. [54] Bo Liu and Yifeng Zhu and Chongkai Gao and Yihao Feng and Qiang Liu and Yuke Zhu and Peter Stone , year =. 2306.03310 , archivePrefix = on PithPith's reviewarXiv
  20. [55] Senyu Fei and Siyin Wang and Junhao Shi and Zihao Dai and Jikun Cai and Pengfang Qian and Li Ji and Xinzhe He and Shiduo Zhang and Zhaoye Fei and Jinlan Fu and Jingjing Gong and Xipeng Qiu , year =. 2510.13626 , archivePrefix = on PithPith's reviewarXiv
show 37 more
  1. [56] Tianxing Chen and Zanxin Chen and Baijun Chen and Zijian Cai and Yibin Liu and Zixuan Li and Qiwei Liang and Xianliang Lin and Yiheng Ge and Zhenyu Gu and Weiliang Deng and Yubin Guo and Tian Nian and Xuanbing Xie and Qiangyu Chen and Kailun Su and Tianling Xu and Guodong Liu ... on Pith
  2. [57] Adrien Bardes and Quentin Garrido and Jean Ponce and Xinlei Chen and Michael Rabbat and Yann LeCun and Mahmoud Assran and Nicolas Ballas , year =. 2404.08471 , archivePrefix = on PithPith's reviewarXiv
  3. [58] Mido Assran and Adrien Bardes and David Fan and Quentin Garrido and Russell Howes and Mojtaba Komeili and Matthew Muckley and Ammar Rizvi and Claire Roberts and Koustuv Sinha and Artem Zholus and Sergio Arnaud and Abha Gejji and Ada Martin and Francois Robert Hogan and Daniel ... on PithPith's reviewarXiv
  4. [59] Lorenzo Mur-Labadia and Matthew Muckley and Amir Bar and Mido Assran and Koustuv Sinha and Mike Rabbat and Yann LeCun and Nicolas Ballas and Adrien Bardes , year =. 2603.14482 , archivePrefix = on PithPith's reviewarXiv
  5. [60] Mahmoud Assran and Quentin Duval and Ishan Misra and Piotr Bojanowski and Pascal Vincent and Michael Rabbat and Yann LeCun and Nicolas Ballas , year =. 2301.08243 , archivePrefix = on PithPith's reviewarXiv
  6. [61] Hyung Won Chung and Le Hou and Shayne Longpre and Barret Zoph and Yi Tay and William Fedus and Yunxuan Li and Xuezhi Wang and Mostafa Dehghani and Siddhartha Brahma and Albert Webson and Shixiang Shane Gu and Zhuyun Dai and Mirac Suzgun and Xinyun Chen and Aakanksha Chowdhery ... on PithPith's reviewarXiv
  7. [62] Cheng Chi and Zhenjia Xu and Siyuan Feng and Eric Cousineau and Yilun Du and Benjamin Burchfiel and Russ Tedrake and Shuran Song , year =. 2303.04137 , archivePrefix = on PithPith's reviewarXiv
  8. [63] Yanjie Ze and Gu Zhang and Kangning Zhang and Chenyuan Hu and Muhan Wang and Huazhe Xu , year =. 2403.03954 , archivePrefix = on PithPith's reviewarXiv
  9. [64] Tony Z. Zhao and Vikash Kumar and Sergey Levine and Chelsea Finn , year =. 2304.13705 , archivePrefix = on PithPith's reviewarXiv
  10. [65] Kevin Black and Noah Brown and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and ... on PithPith's reviewarXiv
  11. [66] Moo Jin Kim and Karl Pertsch and Siddharth Karamcheti and Ted Xiao and Ashwin Balakrishna and Suraj Nair and Rafael Rafailov and Ethan Foster and Grace Lam and Pannag Sanketi and Quan Vuong and Thomas Kollar and Benjamin Burchfiel and Russ Tedrake and Dorsa Sadigh and Sergey L... on PithPith's reviewarXiv
  12. [67] Renming Huang and Chendong Zeng and Wenjing Tang and Jintian Cai and Cewu Lu and Panpan Cai , year =. 2602.08602 , archivePrefix = on PitharXiv
  13. [68] Yingyan Li and Shuyao Shang and Weisong Liu and Bing Zhan and Haochen Wang and Yuqi Wang and Yuntao Chen and Xiaoman Wang and Yasong An and Chufeng Tang and Lu Hou and Lue Fan and Zhaoxiang Zhang , year =. 2510.12796 , archivePrefix = on PithPith's reviewarXiv
  14. [69] Lin Li and Qihang Zhang and Yiming Luo and Shuai Yang and Ruilin Wang and Fei Han and Mingrui Yu and Zelin Gao and Nan Xue and Xing Zhu and Yujun Shen and Yinghao Xu , year =. 2601.21998 , archivePrefix = on PithPith's reviewarXiv
  15. [70] Lishan Yang and Wenxuan Song and Xi Wang and Pingyue Sheng and Zheng Fang and Ziyang Zhou and Junjie He and Haodong Yan and Jiayi Chen and Nan Sun and Qiao Sun and Pengwei Wang and Lingqiao Liu and Yan Wang and Yuxiang Gao and Feras Dayoub and Haoang Li , year =. 2608.08023 , ... on PithPith's reviewarXiv
  16. [71] Haodong Yan and Junfeng Li and Junjie He and Zhide Zhong and MingMing Yu and Wenxuan Song and Jiaguan Zhu and Yangyang Zheng and Yuqiao Du and Jiadi You and Yingjie Cai and Xu Yan and Guanyi Zhao and Bingbing Liu and Haoang Li , year =. 2608.05903 , archivePrefix = on PithPith's reviewarXiv
  17. [72] Yaron Lipman and Ricky T. Q. Chen and Heli Ben-Hamu and Maximilian Nickel and Matt Le , year =. 2210.02747 , archivePrefix = on PithPith's reviewarXiv
  18. [73] William Peebles and Saining Xie , year =. 2212.09748 , archivePrefix = on PithPith's reviewarXiv
  19. [74] Ruijie Zheng and Jing Wang and Scott Reed and Johan Bjorck and Yu Fang and Fengyuan Hu and Joel Jang and Kaushil Kundalia and Zongyu Lin and Loic Magne and Avnish Narayan and You Liang Tan and Guanzhi Wang and Qi Wang and Jiannan Xiang and Yinzhen Xu and Seonghyeon Ye and Jan ... on PithPith's reviewarXiv
  20. [75] Yucheng Hu and Yanjiang Guo and Pengchao Wang and Xiaoyu Chen and. 2024 , eprint = on Pith
  21. [76] Shuang Li and Yihuai Gao and Dorsa Sadigh and Shuran Song , year =. 2503.00200 , archivePrefix = on PithPith's reviewarXiv
  22. [77] Chuning Zhu and Raymond Yu and Siyuan Feng and Benjamin Burchfiel and Paarth Shah and Abhishek Gupta , year =. 2504.02792 , archivePrefix = on PithPith's reviewarXiv
  23. [78] Jingwen Sun and Wenyao Zhang and Zekun Qi and Shaojie Ren and Zezhi Liu and Hanxin Zhu and Guangzhong Sun and Xin Jin and Zhibo Chen , year =. 2602.10098 , archivePrefix = on PitharXiv
  24. [79] Shangchen Miao and Ningya Feng and Jialong Wu and Ye Lin and Xu He and Dong Li and Mingsheng Long , year =. 2602.11832 , archivePrefix = on PitharXiv
  25. [80] Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen-Zhu and Yuanzhi Li and Shean Wang and Lu Wang and Weizhu Chen , year =. 2106.09685 , archivePrefix = on PithPith's reviewarXiv
  26. [81] Quentin Garrido and Randall Balestriero and Laurent Najman and Yann Lecun , year =. 2210.02885 , archivePrefix = on PithPith's reviewarXiv
  27. [82] Mustafa Shukor and Dana Aubakirova and Francesco Capuano and Pepijn Kooijmans and Steven Palma and Adil Zouitine and Michel Aractingi and Caroline Pascal and Martino Russi and Andres Marafioti and Simon Alibert and Matthieu Cord and Thomas Wolf and Remi Cadene , year =. 2506.0... on PithPith's reviewarXiv
  28. [83] Songming Liu and Lingxuan Wu and Bangguo Li and Hengkai Tan and Huayu Chen and Zhengyi Wang and Ke Xu and Hang Su and Jun Zhu , year =. 2410.07864 , archivePrefix = on PithPith's reviewarXiv
  29. [84] Hyung Won Chung and Noah Constant and Xavier Garcia and Adam Roberts and Yi Tay and Sharan Narang and Orhan Firat , year =. 2304.09151 , archivePrefix = on PithPith's reviewarXiv
  30. [85] Aleksandar Vujinovic and Aleksandar Kovacevic , year =. 2501.14622 , archivePrefix = on PithPith's reviewarXiv
  31. [86] Jiangran Lyu and Kai Liu and Xuheng Zhang and Haoran Liao and Yusen Feng and Wenxuan Zhu and Tingrui Shen and Jiayi Chen and Jiazhao Zhang and Yifei Dong and Wenbo Cui and Senmao Qi and Shuo Wang and Yixin Zheng and Mi Yan and Xuesong Shi and Haoran Li and Dongbin Zhao and. 20... on Pith
  32. [87] Jonathan Ho and Tim Salimans , year =. 2207.12598 , archivePrefix = on PithPith's reviewarXiv
  33. [88] Jianlin Su and Yu Lu and Shengfeng Pan and Ahmed Murtadha and Bo Wen and Yunfeng Liu , year =. 2104.09864 , archivePrefix = on PithPith's reviewarXiv
  34. [89] Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success , author=. 2025 , eprint= on Pith
  35. [90] Beyond Future Prediction: Denoising as Generative Adaptation for Robot Control , author =. arXiv preprint arXiv:2609.28339 , year = on PithPith's reviewarXiv
  36. [91] Flex- : A Multi-Stream World-Action Model with Compute Flexibility , author =. arXiv preprint arXiv:2608.10860 , year = on PithPith's reviewarXiv
  37. [92] Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising , author =. Advances in Neural Information Processing Systems , year = on Pith

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.

No comments yet.

Sign in to comment