Pith.

Robotics · arXiv:2610.11480 v1 · submitted 8 Oct 2026 · 86 references

RoboAware: Learning to Coordinate Embodied Skills from Counterfactual Outcomes

Bohan Zhou, Xingbei Chen, Emily Huang, Weilin Ruan, Haojian Huang, Yehang Zhang, Zexi Li, Wenqian Li

show 12 more authors

Qize Yu, Zetian Song, Leyi Wu, Jinghao Li, Mingxuan Song, Xinrun Xu, Zongyang Qiu, Yangkai Wei, Tianyi Zhang, Kaiwen Zhou, Yinchuan Li, James Cheng

The paper says

“Embodied coding agents can combine modular robot skills with frozen end-to-end policies, yet effective composition requires anticipating which policy family will succeed in the current physical state.”

more of the abstract

We present RoboAware, which builds on coding agents' skill orchestration by learning only a state-conditioned responsibility coordinator from counterfactual outcomes. Inspired by the success of REPL, we propose the P5 schema and formulate a hierarchical MDP based on it. P5 organizes skills uniformly into five semantic stages, defining where responsibility can be compared. To address the lack of counterfactual branch outcomes in existing work, we introduce State-Locked Counterfactual Branching (SCB), which restores the same training state to generate and execute a code block from each admissible family, exposing outcomes that selected-branch experience leaves unobserved. Building on this, we propose Execution-Aware Learning (EAL), which combines Monte Carlo tree search with Q-learning to distill these outcomes into family-conditioned values. At deployment, the coordinator selects the policy family according to observable context, and the frozen coding agent generates the next local code block. Comprehensive single-episode evaluations on 100 tasks show that RoboAware reaches a 77.0% overall success rate, with SOTA averages of 90.0% on RoboSuite, 73.8% on diverse LIBERO-Pro task clusters, and 90.0% on challenging RoboTwin bimanual tasks, outperforming existing code-as-policy and VLA-harness baselines.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 10 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. Novel SCB+EAL method for learning policy-family coordination is well-designed, but the simulator-gated training data and lack of OOD handling in the offline Q-learning limit external validity. This line is wrong

3 major objections · 5 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 86 references say.

The paper says, at Algorithm 1, §2.2.2

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

During expansion (lines 11–13), each admissible family is evaluated with a single code generation and execution. End-to-end policies (e.g., π0.5, FastWAM) are stochastic; a single outcome may be a poor estimate of the expected return. Although later UCB visits can re-evaluate selected families, unselected families contribute only their initial sample to the DSCB dataset used for Q-learning.

more of objection 1

The paper does not analyze the sensitivity of Qθ to this single-sample noise or provide evidence (e.g., repeatability experiments) that the Q-values stabilize. Without such analysis, the reliability of the coordinator in states where one family’s outcome happens to be an outlier is unclear.

The paper says, at Equation (4), §2.2.2

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The Q-learning objective in Eq. (4) uses a standard Bellman update on the offline dataset DSCB, but does not include any conservative regularization (e.g., CQL) to address distributional shift between the data-collection policy (MCTS-UCB) and the target policy (argmax Qθ).

more of objection 2

When certain state-family pairs are underrepresented in DSCB, the max operation may extrapolate and overestimate Q-values. The paper should at minimum calibrate the learned Q-values (e.g., by comparing predicted and empirical returns on held-out states) or discuss why extrapolation is not a concern given the data-collection protocol.

The paper says, at §3.2, Table 3

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The comparison against Harness-VLA is central to the paper's claims, yet the Harness-VLA results in Table 3 use a single-seed evaluation protocol. Appendix D.1 reports that Harness-VLA achieves 56.6% valid success rate on 80 tasks (seeds 41–43), but the main table shows 50.0% overall on 100 tasks (seeds 1–10).

more of objection 3

The discrepancy is not clearly explained, and the main table omits the execution-coverage statistics reported in Appendix D.1. The authors should reconcile these numbers or report coverage-adjusted success rates to ensure a fair comparison.

5 minor objections

The paper says, at Table 4

Pith says: minor objection 1 of 5

The group labels (A, B, C) are described in the text but the table itself does not show these letters; adding them explicitly would improve readability.

The paper says, at Figure 5

Pith says: minor objection 2 of 5

The y-axis label reads 'Average RoboSuite success rates (%) across budgets K ∈ {5, 10, 20, 50}' but the caption says 'Adopted72.0%' in the first column—this seems to be a formatting artifact; please clean up the caption text.

show 3 more

The paper says, at §2.2.2

Pith says: minor objection 3 of 5

The symbol ξi is defined as a snapshot containing privileged information (oracle masks, poses, reward evaluator). At first read it is unclear how ξi differs from xi; the distinction between observable context and privileged state should be clarified earlier in that paragraph.

The paper says, at Appendix A.2

Pith says: minor objection 4 of 5

The 'neighboring-state sampling' procedure draws S=4 roots per stage entry within a time window Δ. The paper should state whether these roots share overlapping transitions in DSCB and whether the group-level train/validation split (§A.4) fully prevents same-neighborhood leakage; the current description is clear but a brief justification of the 90/10 ratio would help.

The paper says, at Table 12

Pith says: minor objection 5 of 5

Hyperparameters are listed but the translational threshold ϵ_p = 0.02 m and rotational threshold ϵ_R = π/12 rad are not motivated; a short justification (e.g., based on task geometry or prior work) would strengthen the reproducibility.

What would settle it

随机选取10个训练任务中的P5阶段入口状态,对每个状态重复10次执行SCB中的E2E代码块(不同随机种子),计算每个状态的成功率标准差。如果标准差均值>0.15,说明单次反事实比较噪声过大。同时,从DSCB中抽取100个转移,用蒙特卡洛Rollout(10次平均)估计真实期望回报,与Qθ预测值比较:若平均绝对误差>0.2,则表明离线Q值存在显著偏差。

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 10 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 55 s · 37,974 tokens read, 5,072 written
  3. Editor wrote the verdict above 13 s · 38,947 tokens read, 1,413 written
  4. Referee major revision: 3 major objections, 5 minor 26 s · 38,332 tokens read, 2,483 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

The paper introduces a clear, well-motivated idea: learn a state-conditioned coordinator that picks between modular and end-to-end policy families at each P5 skill stage, using counterfactual simulator data. The combination of State-Locked Counterfactual Branching (SCB) and Execution-Aware Learning (EAL)—wrapping MCTS + Q-learning around same-state paired rollouts—is genuinely new and well executed. The ablations are thorough and honestly reported; the 77% overall success rate across three benchmarks is solid and beats sensible baselines. The authors also state their limitations explicitly (simulator oracle dependence, coarse interruption granularity), which is refreshing.

However, the load-bearing assumption is that training has access to a simulator that can save/restore exact physical states and emit oracle masks, poses, and terminal rewards. The authors acknowledge this is an open problem, but it means the method cannot currently be applied to real-world tasks or tasks that aren't faithfully simulable. This limits the practical reach of the result.

The stress-test note raises a fair point about single-sample variance in SCB: at a new node only one execution per family is used, and for stochastic policies this could produce noisy Q targets. However, the MCTS loop revisits family edges via UCB selection and the budget of K=20 means multiple samples over time, which mitigates the bias to some degree. The offline Q-learning without any conservative regularization (like CQL) is a real concern: missing data for some state-action pairs could lead to extrapolation errors. The paper provides no calibration evidence for Q-values. These are genuine soft spots that weaken the claim about the coordinator's reliability in unexplored states.

On balance: the paper deserves a serious peer review. The core idea is clever and the engineering is impressive. The main revisions needed are (1) release code and data so others can verify, (2) control for the privileged training signal when comparing against baselines that lack it, and (3) address the off-policy evaluation issue—even a simple analysis of Q-value error or using a CQL-style penalty would help.

Recommendation: engage with the work, but hold the revision bar relatively high on reproducibility and OOD robustness.

The strongest objection

论文的核心是学习一个状态条件化的协调器,通过SCB收集的成对反事实结果来训练Q函数。然而,SCB在每个新扩展节点仅对每个家族采样并执行一次(Algorithm 1第11-13行)。端到端策略(如π0.5,FastWAM)通常是随机的(例如扩散去噪过程),单次执行结果可能不能代表该家族的期望回报。虽然MCTS后续访问可能重新评估同一家族(第10行),但这仅对UCB选择的家族进行,且首次扩展的单次采样直接进入DSCB作为Q学习的目标。如果端到端策略的方差较大,单次反事实比较会产生有偏的Q值目标,进而误导家族选择。此外,DSCB是离线收集的固定数据集,Q学习(Eq. 4)使用Bellman更新,但未引入任何保守性正则化(如CQL)来处理行为策略(MCTS-UCB)与目标策略(argmax Q)之间的分布偏移。当数据未覆盖某些状态-动作对时,max操作可能外推并高估Q值。论文没有提供Q值标定证据或分布外鲁棒性分析。这两个问题组合起来,可能削弱协调器在未充分探索状态下的决策可靠性,直接影响77.0%总成功率的可重复性。

读者指出最薄弱的假设是训练依赖模拟器保存/恢复和oracle信号,这是一个外部有效性限制。我提出的concern聚焦于方法内部的技术可靠性:SCB对随机性策略的单次采样可能导致有噪目标,且离线Q学习缺乏保守处理可能引入分布外高估。这两个问题都是论文未明确讨论的潜在缺陷,且直接关联到Q值学习和协调器选择的核心。读者尚未强调它们,因此属于补充性质的薄弱点。尽管如此,这些concern可能通过更大数据量或随机种子平均在实践中减轻,且现有实验效果良好,不影响整体CONDITIONAL判断,故verdict不变。

The referee's report

If the results hold, RoboAware provides a principled method for heterogeneous policy orchestration by leveraging simulator-based counterfactual outcomes to learn a responsibility coordinator, without modifying the coding agent or low-level executors. The P5 skill schema offers a structured decomposition that enables fine-grained family selection. The reported gains over strong baselines (e.g., Harness-VLA, ASPIRE) suggest that explicit same-state counterfactuals can improve deployment decisions. The approach is reproducible through the detailed algorithmic description and ablations, and the use of disjoint training/evaluation seeds supports generalization claims.

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

The paper argues that embodied coding agents fail to anticipate which style of skill execution—modular API composition or a frozen end-to-end policy—will succeed in the current physical state. RoboAware addresses this by training only a lightweight coordinator that selects the better policy family at each of five standardized skill stages. To generate the training signal, the method repeatedly restores the simulator to the exact same state, executes both families from that state, and records their outcomes. A Q-function learned from these paired outcomes then guides family selection at test time without needing a simulator. The claim is that this makes anticipated physical outcomes an explicit basis for composing complementary frozen skills, yielding a 77.0% overall success rate on 100 manipulation tasks across three benchmarks.

What carries the argument. The P5 skill schema decomposes every embodied skill into five consecutive stages—Perceive, Propose, Pre-manipulate, Perform, Post-manipulate—creating uniform decision points where a policy family must be chosen. State-Locked Counterfactual Branching (SCB) is the training-time operator that restores the simulator to the exact same physical state and executes a code block from each admissible family (Modular vs End-to-End), yielding paired outcomes that would otherwise be hidden. Execution-Aware Learning (EAL) combines Monte Carlo tree search with UCB-guided node expansion to explore these paired outcomes, then distills them into a Q-function Qθ(xi, c) via Bellman updates on the collected transitions. At test time, the frozen coordinator selects the family with the highest Q-value, and the coding agent generates the corresponding code block without any simulator access or online search.

Core claim. RoboAware establishes that a state-conditioned responsibility coordinator, trained exclusively on counterfactual simulator data (State-Locked Counterfactual Branching) and learned via a combination of Monte Carlo tree search and Q-learning (Execution-Aware Learning), can reliably decide at each skill stage whether to use a modular code block or an end-to-end policy block. The P5 skill schema (Perceive, Propose, Pre-manipulate, Perform, Post-manipulate) provides the uniform semantic boundaries where this decision is made. The paper demonstrates that neither family alone suffices, that paired same-state outcomes improve learning over observed-only trajectories, and that per-stage coordination outperforms fixed handoff or single-episode choices. On 100 held-out tasks across LIBERO-Pro, RoboSuite, and RoboTwin 2.0, the method achieves a 77.0% overall success rate, outperforming existing code-as-policy, end-to-end policy, and VLA-harness baselines.

The premise it rests on. Training requires a simulator that can save and restore exact physical states (including object poses, mask oracles, and terminal reward evaluators) and can execute both the modular and end-to-end families from the same restored state; this is acknowledged as an open problem for removal.

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

If RoboAware's approach generalizes beyond the three evaluated benchmarks, it offers a principled way to combine any set of complementary policy families (modular, end-to-end, retrieval-based, reinforcement-learned) by learning when each is preferable from counterfactual data.

The P5 schema could serve as a standard interface for skill decomposition in hierarchical robot control, making policy-family coordination a tractable learning problem rather than a hand-crafted switching rule.

Because only the lightweight coordinator is learned while the coding agent and API library remain frozen, RoboAware can incorporate improved modular backends or new end-to-end policies without retraining the coordinator—only the counterfactual data need to be recollected.

The method's reliance on simulator save-restore and oracle signals during training suggests that the approach is most directly applicable in simulation-first training pipelines; transferring to real-world settings would require either a reliable simulator or approximate counterfactual generation.

RoboAware's framework implicitly assumes that the choice between policy families is a Markov decision at each P5 boundary, but real-world skill execution may have long-range dependencies (e.g., a poor grasp choice propagates failure through later stages). The learned Q-function may need to capture these dependencies beyond the immediate next stage.

The paper's sensitivity analysis (Takeaway 5) shows small performance gaps across three coding-agent backends and four robot embodiments, suggesting that the learned coordinator is robust to changes in the code generator and hardware—a desirable property for practical deployment.

The open problem of finer-grained interruption during an atomic end-to-end call (acknowledged in Section 5) hints that the current P5-boundary-only switching may be too coarse for tasks where transient failures within a stage could be corrected by switching families mid-stage.

The 37.7% of failures categorized as 'Others' (insufficient trace detail) indicates that the current evaluation infrastructure may obscure important failure patterns; future work should improve failure categorization to guide further improvements.

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

Embodied coding agents can combine modular robot skills with frozen end-to-end policies, yet effective composition requires anticipating which policy family will succeed in the current physical state. We present RoboAware, which builds on coding agents' skill orchestration by learning only a state-conditioned responsibility coordinator from counterfactual outcomes. Inspired by the success of REPL, we propose the P5 schema and formulate a hierarchical MDP based on it. P5 organizes skills uniformly into five semantic stages, defining where responsibility can be compared. To address the lack of counterfactual branch outcomes in existing work, we introduce State-Locked Counterfactual Branching (SCB), which restores the same training state to generate and execute a code block from each admissible family, exposing outcomes that selected-branch experience leaves unobserved. Building on this, we propose Execution-Aware Learning (EAL), which combines Monte Carlo tree search with Q-learning to distill these outcomes into family-conditioned values. At deployment, the coordinator selects the policy family according to observable context, and the frozen coding agent generates the next local code block. Comprehensive single-episode evaluations on 100 tasks show that RoboAware reaches a 77.0% overall success rate, with SOTA averages of 90.0% on RoboSuite, 73.8% on diverse LIBERO-Pro task clusters, and 90.0% on challenging RoboTwin bimanual tasks, outperforming existing code-as-policy and VLA-harness baselines.

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 · 2026-10-09

The authors' main claim, as the classifier picked it out 0.99

“Comprehensive single-episode evaluations on 100 tasks show that RoboAware reaches a 77.0% overall success rate, with SOTA averages of 90.0% on RoboSuite, 73.8% on diverse LIBERO-Pro task clusters, and 90.0% on challenging RoboTwin bimanual tasks, outperforming existing code-as-policy and VLA-harness baselines.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.10 · states limitation 0.13 · human participants 0.03 · controlled comparison 0.90 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.97 · 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.

SentenceRoleThe authors' own workStates the main conclusionStrength
Embodied coding agents can combine modular robot skills with frozen end-to-end policies, yet effective composition requires anticipating which policy family will succeed in the current physical state. background 0.93 0.06 0.04 asserted 0.98
We present RoboAware, which builds on coding agents' skill orchestration by learning only a state-conditioned responsibility coordinator from counterfactual outcomes. contribution 0.98 0.87 0.12 asserted 1.00
Inspired by the success of REPL, we propose the $P^5$ schema and formulate a hierarchical MDP based on it. method 0.70 0.80 0.02 asserted 0.93
$P^5$ organizes skills uniformly into five semantic stages, defining where responsibility can be compared. method 0.89 0.77 0.03 asserted 1.00
To address the lack of counterfactual branch outcomes in existing work, we introduce State-Locked Counterfactual Branching (SCB), which restores the same training state to generate and execute a code block from each admissible family, exposing outcomes that selected-branch experience leaves unobserved. method 0.56 0.90 0.03 asserted 1.00
Building on this, we propose Execution-Aware Learning (EAL), which combines Monte Carlo tree search with Q-learning to distill these outcomes into family-conditioned values. method 0.89 0.91 0.04 asserted 0.97
At deployment, the coordinator selects the policy family according to observable context, and the frozen coding agent generates the next local code block. method 0.99 0.53 0.04 asserted 0.99
Comprehensive single-episode evaluations on 100 tasks show that RoboAware reaches a 77.0% overall success rate, with SOTA averages of 90.0% on RoboSuite, 73.8% on diverse LIBERO-Pro task clusters, and 90.0% on challenging RoboTwin bimanual tasks, outperforming existing code-as-policy and VLA-harness baselines. result 1.00 0.96 0.81 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.80

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, 2026-10-09: 1 request, 6,530 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. Overview of RoboAware. (a) The P 5 skill schema decomposes each skill into five stages—perceive, propose, pre-manipulate, perform, and post-manipulate—and unifies primitive skills (e.g., pick, place, push, pull) and compositional skills (e.g., handover) under one representation. (b) A hierarchical MDP selects a policy family ci ∈ {Mod,E2E} at each P 5 stage decision boundary; a coding agent then generates each corresponding code block ai. (c) EAL learns family-conditioned coordination: SCB evaluates modular and end-to-end branches from the same restored state into DSCB, Q-learning…
Figure 2 of the paper
Figure 2. Method overview of EAL in RoboAware. (1) State-locked counterfactual search repeats Selection (UCB), Expansion (SCB paired Mod/E2E branches from a restored ξi), Rollout (suffix return Gi), and Backup within budget K, retaining successful, failed, and nonpreferred transitions in DSCB. (2) A VLM backbone with an MLP head maps context–family tokens ending in <VALUE> to Qθ(xi, ci), trained by Bellman updates on skill-completion and terminal-task feedback. (3) At deployment, frozen Qθ induces ρθ and selects c ∗ i = arg maxc ρθ(c | xi); a frozen coding agent generates a local code block…
Figure 4 of the paper
Figure 4. A representative LIBERO-Pro task compar￾ison and visualization of the responsibility preference distribution ρθ for RoboAware, RoboAware-EO, and RoboAware-MO. Takeaway 1: Neither policy family alone provides broad high-success task coverage. We restrict code generation to modular composition (RoboAware-MO) or end-to-end blocks (RoboAware-EO), keeping the same API library and tasks as RoboAware. Success rate (SR) reductions in group A of
Figure 6 of the paper
Figure 6. RoboTwin 2.0 success rates (%) on a four-task subset across coding-agent backends (rows) and em￾bodiments (columns). Color bar shows the SR range. Takeaway 4: UCB-guided exploration uses a limited branching budget more effectively. To test how a finite simulation budget is allocated during branching, we compare RoboAware with RoboAware-UA, which allocates evaluations uniformly over admissible families instead of the UCB rule in Equation (1)

Showing 4 of the paper's 6 figures. Figures are the authors', reproduced from the paper.

References (86)
  1. [1] Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. InRobotics: Science and Systems (RSS), 2023 on Pith
  2. [2] 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
  3. [3] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3D diffusion policy: Generalizable visuomotor policy learning via simple 3D representations. InRobotics: Science and Systems (RSS), 2024 on Pith
  4. [4] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, et al. RT-1: Robotics transformer for real-world control at scale. InRobotics: Science and Systems (RSS), 2023 on Pith
  5. [5] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. InConference on Robot Learning (CoRL), 2023 on Pith
  6. [6] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, et al. OpenVLA: An open-source vision-language-action model. InConference on Robot Learning (CoRL), 2024 on Pith
  7. [7] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Robert Equi, et al.π0: A vision-language-action flow model for general robot control. InRobotics: Science and Systems (RSS), 2025 on Pith
  8. [8] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. InIEEE International Conference on Robotics and Automation (ICRA), 2023 on Pith
  9. [9] Letian Fu, Justin Yu, Karim El-Refai, Ethan Kou, Haoru Xue, et al. CaP-X: A framework for benchmarking and improving coding agents for robot manipulation.arXiv preprint arXiv:2603.22435, 2026 on PithPith's reviewarXiv
  10. [10] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Fei-Fei Li. VoxPoser: Composable 3D value maps for robotic manipulation with language models. InConference on Robot Learning (CoRL), 2023 on Pith
  11. [11] Yixian Zhang, Huanming Zhang, Feng Gao, Xiao Li, Zhihao Liu, et al. Harness VLA: Steering frozen VLAs into reliable manipulation primitives via memory-guided agents.arXiv preprint arXiv:2607.08448, 2026 on PithPith's reviewarXiv
  12. [12] Jiaheng Hu, Mohit Shridhar, Caden Lu, Dhruv Shah, Hao-Tien Lewis Chiang, Jie Tan, and Annie Xie. What matters in orchestrating robot policies: A systematic study of hierarchical VLA agents.arXiv preprint arXiv:2606.10267, 2026 on PithPith's reviewarXiv
  13. [13] Jinbang Huang, Yuanzhao Hu, Zhiyuan Li, Ran Qi, Yixin Xiao, et al. RoboHarness: Memory-driven orchestration of heterogeneous robot policies for long-horizon planning.arXiv preprint arXiv:2607.18060, 2026 on PithPith's reviewarXiv
  14. [14] Anthony Z. Liu, Xinhe Wang, Jacob Sansom, Yao Fu, Jongwook Choi, Sungryull Sohn, Jaekyeom Kim, and Honglak Lee. Interactive and expressive code-augmented planning with large language models. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguist... on Pithdoi:10.18653/v1/2025.acl-long.994
  15. [15] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating situated robot task plans using large language models. InIEEE International Conference on Robotics and Automation (ICRA), 2023 on Pith
  16. [16] Jiachen Zhang, Junnan Nie, Junyi Lao, Wei Cheng, Chenghao Liu, Jiaxin Jiang, and Songfang Huang. What frozen VLAs already know about success: A probing study of value-like structure in foundation robot policies. arXiv preprint arXiv:2605.28527, 2026 on PithPith's reviewarXiv
  17. [17] Yuqian Yuan, Wentong Li, Zhaocheng Li, Yutong Lin, Juncheng Li, Siliang Tang, Jun Xiao, Yueting Zhuang, and Wenqiao Zhang. InstructSAM: Segment any instance with any instructions.arXiv preprint arXiv:2605.26102, 2026 on PithPith's reviewarXiv
  18. [18] Snehal Jauhri, Ishikaa Lunawat, and Georgia Chalvatzaki. Learning any-view 6DoF robotic grasping in cluttered scenes via neural surface rendering. InRobotics: Science and Systems (RSS), 2024 on Pith
  19. [19] Balakumar Sundaralingam, Adithyavairavan Murali, and Stan Birchfield. cuRoboV2: Dynamics-aware motion generation with depth-fused distance fields for high-DoF robots.arXiv preprint arXiv:2603.05493, 2026 on PithPith's reviewarXiv
  20. [20] Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, et al.π0.5: A vision-language-action model with open-world generalization. InProceedings of The 9th Conference on Robot Learning, volume 305 ofProceedings of Machine Learning Researc... on Pith
show 66 more
  1. [21] Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-WAM: Do world action models need test-time future imagination?arXiv preprint arXiv:2603.16666, 2026. URL https://arxiv.org/abs/2603.16666 on PithPith's reviewarXiv
  2. [22] Xueyang Zhou, Yangming Xu, Guiyao Tie, Yongchao Chen, Guowen Zhang, Duanfeng Chu, Pan Zhou, and Lichao Sun. LIBERO-PRO: Towards robust and fair evaluation of vision-language-action models beyond memorization.arXiv preprint arXiv:2510.03827, 2025 on PithPith's reviewarXiv
  3. [23] Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martín-Martín, Abhishek Joshi, Kevin Lin, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning.arXiv preprint arXiv:2009.12293, 2020. 10 RoboAware: Learning to Coordinate Em... on PithPith's reviewarXiv
  4. [24] Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Qiwei Liang, Zixuan Li, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et al. RoboTwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation. InInternational ... on Pith
  5. [25] Weiheng Zhao, Haoyi Jiang, Xin Shi, Liu Liu, Fan Huang, Zhizhong Su, Wei Sui, and Xinggang Wang. Faster-WAM: Efficient inference-time future conditioning for robust world action models.arXiv preprint arXiv:2608.04404, 2026. URL https://arxiv.org/abs/2608.04404 on PithPith's reviewarXiv
  6. [26] Junyi Zhang, Jiaxin Ge, Hanjun Yoo, Letian Fu, Zihan Yang, Yaowei Liu, Raj Saravanan, Shaofeng Yin, Justin Yu, Dantong Niu, et al. Playful agentic robot learning.arXiv preprint arXiv:2606.19419, 2026 on PithPith's reviewarXiv
  7. [27] Runyu Lu, Yubo Wu, Ethan Kou, Letian Fu, Wenli Xiao, Ajay Mandlekar, Yinzhen Xu, Guanya Shi, Ken Goldberg, Ang Chen, Mosharaf Chowdhury, Yuke Zhu, Linxi “Jim” Fan, and Guanzhi Wang. ASPIRE: Agentic /skills discovery for robotics.arXiv preprint arXiv:2607.00272, 2026 on PithPith's reviewarXiv
  8. [28] Liane Galanti, Dhruv Shah, and Tri Dao. Addressing the orchestration gap in generalist robots via physical agency.arXiv preprint arXiv:2607.21725, 2026 on PithPith's reviewarXiv
  9. [29] Shuo Liu, Ishneet Sukhvinder Singh, Yiqing Xu, Jiafei Duan, and Ranjay Krishna. VLS: Steering pretrained robot policies via vision-language models.arXiv preprint arXiv:2602.03973, 2026 on PitharXiv
  10. [30] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. LLM-Planner: Few-shot grounded planning for embodied agents with large language models. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2998–3009, 20... on Pithdoi:10.1109/iccv51070.2023.00280
  11. [31] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, et al. Do as I can, not as I say: Grounding language in robotic affordances. InProceedings of The 6th Conference on Robot Learning, volume 205 ofProceedings of Machine Learning Research, pages 287–318, 2023... on Pith
  12. [32] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, et al. Inner monologue: Embodied reasoning through planning with language models. InConference on Robot Learning (CoRL), 2023 on Pith
  13. [33] Yao Mu, Junting Chen, Qing-Long Zhang, Shoufa Chen, Qiaojun Yu, Chongjian Ge, Runjian Chen, Zhixuan Liang, Mengkang Hu, Chaofan Tao, Peize Sun, Haibao Yu, Chao Yang, Wenqi Shao, Wenhai Wang, Jifeng Dai, Yu Qiao, Mingyu Ding, and Ping Luo. RoboCodeX: Multimodal code generation ... on Pith
  14. [34] Guang Yin, Yitong Li, Yixuan Wang, Dale McConachie, Paarth Shah, Kunimatsu Hashimoto, Huan Zhang, Katherine Liu, and Yunzhu Li. CodeDiffuser: Attention-enhanced diffusion policy via VLM-generated code for instruction ambiguity. InRobotics: Science and Systems (RSS), 2025 on Pith
  15. [35] Vaishak Kumar. Act-observe-rewrite: Multimodal coding agents as in-context policy learners for robot manipulation.arXiv preprint arXiv:2603.04466, 2026 on PitharXiv
  16. [36] Karim Elmaaroufi, Justin Svegliato, Sarunas Kalade, Graham Schelle, Sanjit A. Seshia, and Matei Zaharia. RHO: Your coding agent is secretly a roboticist.arXiv preprint arXiv:2606.16458, 2026 on PitharXiv
  17. [37] Kaiyuan Chen, Shuangyu Xie, Letian Fu, Justin Yu, William Pacini, Sandeep Bajamahal, Hudson Kim, Jaimyn Drake, Daehwa Kim, Haoru Xue, Jonathan Francis, Christian Juette, Peter Schaldenbrand, Muhammet Yunus Seker, Ruwan Wickramarachchi, Uksang Yoo, Guanzhi Wang, Adithyavairavan... on PithPith's reviewarXiv
  18. [38] Dhia Naouali, Minghan Wu, Claudia Wong, Abhinav Puthran, and Omar G. Younis. VLCP: Vision language control policy closed-loop code replanning for robot manipulation.arXiv preprint arXiv:2608.16978, 2026 on PithPith's reviewarXiv
  19. [39] Chen-Yu Lin, Jing-Wen Chen, Hsueh-En Chang, Hung-An Chen, Sheng-Hsun Chang, Chi-Pin Huang, Fu-En Yang, Min-Hung Chen, Yi-Ting Chen, Yu-Chiang Frank Wang, and Shao-Hua Sun. PhysCaP: Grounding code-as-policy agent with physics-informed exploration.arXiv preprint arXiv:2608.21031, 2026 on PithPith's reviewarXiv
  20. [40] Dibya Ghosh, Homer Rich Walke, Karl Pertsch, Kevin Black, Oier Mees, et al. Octo: An open-source generalist robot policy. InRobotics: Science and Systems (RSS), 2024 on Pith
  21. [41] Zixing Lei, Changxing Liu, Yichen Xiong, Minhao Xiong, Yuanzhuo Ding, Zhipeng Zhang, Weixin Li, and Siheng Chen. Towards long-horizon embodied agents with tool-aligned vision-language-action models.arXiv preprint arXiv:2605.13119, 2026 on PithPith's reviewarXiv
  22. [42] Khoa Vo, Sieu Tran, Taisei Hanyu, Yuki Ikebe, Duy Nguyen, et al. CodeGraphVLP: Code-as-planner meets semantic-graph state for non-markovian vision-language-action models.arXiv preprint arXiv:2604.22238, 2026 on PithPith's reviewarXiv
  23. [43] Jiaqi Peng, Xiqian Yu, Delin Feng, Yuqiang Yang, Wenzhe Cai, et al. Cortex: A bidirectionally aligned embodied agent framework for long-horizon manipulation.arXiv preprint arXiv:2607.05377, 2026 on PithPith's reviewarXiv
  24. [44] Siyi Chen, Hugo Hadfield, Alex Zook, Mikaela Angelina Uy, Chan Hee Song, Erwin Coumans, Xuning Yang, Faisal Ladhak, Qing Qu, Stan Birchfield, Jonathan Tremblay, and Valts Blukis. VoLo: A physical orchestrator 11 RoboAware: Learning to Coordinate Embodied Skills from Counterfac... on PithPith's reviewarXiv
  25. [45] Sihyung Yoon, Minjong Yoo, Sanghyun Ahn, Seojeong Choi, and Honguk Woo. RoboBRIDGE: A modular framework for bridging policies to robust real-world robotic agents. InIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026 on Pith
  26. [46] Yiteng Chen, Zhe Cao, Hongjia Ren, Chenjie Yang, Wenbo Li, et al. RoboRouter: Training-free policy routing for robotic manipulation.arXiv preprint arXiv:2603.07892, 2026 on PithPith's reviewarXiv
  27. [47] Xingyu Ren, Chugang Yi, and Youran Sun. RouterVLA: Budgeted commissioning and expert onboarding for growing VLA pools.arXiv preprint arXiv:2606.27355, 2026 on PithPith's reviewarXiv
  28. [48] Yinlong Dai, Andre Keyser, and Dylan P. Losey. Prepare before you act: Learning from humans to rearrange initial states.arXiv preprint arXiv:2509.18043, 2025 on PitharXiv
  29. [49] Meng Li, Zhen Zhao, Zhengping Che, Fei Liao, Kun Wu, et al. SwitchVLA: Execution-aware task switching for vision-language-action models.arXiv preprint arXiv:2506.03574, 2025 on PithPith's reviewarXiv
  30. [50] Ke Rui, Yushen Zuo, Jiawei Wang, Haoran Jia, Jinming Ma, Weitao Zhou, and Minglei Li. Diagnosing semantic handoff failures in agent-orchestrated vision-language-action skill composition. InRSS SemRob Workshop, 2026 on Pith
  31. [51] Ming Shang, Yuchen Huang, Jiaoyang Chen, Haoyuan Hu, Han Yu, Liping Song, Luyun Feng, Shuo Bao, Wei Dong, Xinzhou Wang, and Fuchun Sun. BICPO-VLA: Behavior-identified continuation preference optimization for smooth asynchronous vision-language-action control.arXiv preprint arX... on PithPith's reviewarXiv
  32. [52] Carl Qi, Xiaojie Wang, Silong Yong, Stephen Sheng, Huitan Mao, et al. Self-refining vision language model for robotic failure detection and reasoning. InInternational Conference on Learning Representations (ICLR), 2026 on Pith
  33. [53] Prasun Saurabh, Pablo Valle, Aitor Arrieta, Shaukat Ali, and Paolo Arcaini. VISOR: A vision-language model-based test oracle for testing robots. InProceedings of the 3rd ACM International Conference on AI-Powered Software, pages 197–207, 2026. doi: 10.1145/3805760.3814910 on Pithdoi:10.1145/3805760.3814910
  34. [54] Shirui Chen, Cole Harrison, Ying-Chun Lee, Angela Jin Yang, Zhongzheng Ren, et al. TOPReward: Token probabilities as hidden zero-shot rewards for robotics.arXiv preprint arXiv:2602.19313, 2026 on PithPith's reviewarXiv
  35. [55] Yanru Wu, Weiduo Yuan, Ang Qi, Vitor Guizilini, Jiageng Mao, and Yue Wang. Large reward models: Generalizable online robot reward generation with vision-language models.arXiv preprint arXiv:2603.16065, 2026 on PithPith's reviewarXiv
  36. [56] Jacky Kwok, Christopher Agia, Rohan Sinha, Matt Foutter, Shulu Li, Ion Stoica, Azalia Mirhoseini, and Marco Pavone. RoboMonkey: Scaling test-time sampling and verification for vision-language-action models. In Proceedings of The 9th Conference on Robot Learning, volume 305 ofP... on Pith
  37. [57] Dayou Li, Jiuzhou Lei, Hao Wang, Lulin Liu, Yunhao Yang, et al. Learning actionable manipulation recovery via counterfactual failure synthesis.arXiv preprint arXiv:2603.13528, 2026 on PitharXiv
  38. [58] Catherine Glossop, William Chen, Arjun Bhorkar, Dhruv Shah, and Sergey Levine. CAST: Counterfactual labels improve instruction following in vision-language-action models.IEEE Robotics and Automation Letters, 11(10): 11785–11792, 2026. doi: 10.1109/LRA.2026.3726383 on Pithdoi:10.1109/lra.2026.3726383
  39. [59] Haru Kondoh, Kei Ota, Asako Kanezaki, and Yueh-Hua Wu. CounterAlign: Counterfactual supervision for vision-language-action models.arXiv preprint arXiv:2608.21740, 2026 on PithPith's reviewarXiv
  40. [60] Xin Ding, Liang Mi, Mingzhe Huang, Zixuan Wang, Chao Zhang, Zixu Hao, Fu Chen, Xiangyu Li, Yikai Zheng, Yaoyu Guo, Weijun Wang, Kun Li, Hao Wu, Yunxin Liu, and Ting Cao. Zettaζ: An efficient closed-loop embodied harness for self-evolving physical intelligence.arXiv preprint ar... on PithPith's reviewarXiv
  41. [61] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, et al. AnyGrasp: Robust and efficient grasp perception in spatial and temporal domains.IEEE Transactions on Robotics, 39(5):3929–3945, on Pith
  42. [62] Adithyavairavan Murali, Balakumar Sundaralingam, Yu-Wei Chao, Wentao Yuan, Jun Yamada, Mark Carlson, et al. GraspGen: A diffusion-based framework for 6-DOF grasping with on-generator training. InIEEE International Conference on Robotics and Automation (ICRA), 2026 on Pith
  43. [63] Xiao-Ming Wu, Jia-Feng Cai, Jian-Jian Jiang, Dian Zheng, Yi-Lin Wei, and Wei-Shi Zheng. An economic framework for 6-DoF grasp detection. InEuropean Conference on Computer Vision (ECCV), pages 357–375, on Pith
  44. [64] Siang Chen, Pengwei Xie, Wei Tang, Dingchang Hu, Yixiang Dai, and Guijin Wang. Region-aware grasp framework with normalized grasp space for efficient 6-DoF grasping. InProceedings of The 8th Conference on Robot Learning, volume 270 ofProceedings of Machine Learning Research, p... on Pith
  45. [65] Jiahui Yang, Jason Jingzhou Liu, Yulong Li, Youssef Khaky, Kenneth Shaw, and Deepak Pathak. Deep reactive policy: Learning reactive manipulator motion planning for dynamic environments. InProceedings of The 9th Conference on Robot Learning, volume 305 ofProceedings of Machine ... on Pith
  46. [66] Wil Thomason, Zachary Kingston, and Lydia E. Kavraki. Motions in microseconds via vectorized sampling-based planning. InIEEE International Conference on Robotics and Automation (ICRA), pages 12 RoboAware: Learning to Coordinate Embodied Skills from Counterfactual Outcomes 8749... on Pithdoi:10.1109/icra57147.2024.10611190
  47. [67] Adam Fishman, Aaron Walsman, Mohak Bhardwaj, Wentao Yuan, Balakumar Sundaralingam, Byron Boots, and Dieter Fox. Avoid everything: Model-free collision avoidance with expert-guided fine-tuning. InProceedings of The 8th Conference on Robot Learning, volume 270 ofProceedings of M... on Pith
  48. [68] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Optimizing speed and success. InRobotics: Science and Systems (RSS), 2025. doi: 10.15607/RSS.2025.XXI.017. URL https://www.roboticsproceedings.org/rss21/p017.html on Pithdoi:10.15607/rss.2025.xxi.017
  49. [69] Yang Zhang, Jiangyuan Zhao, Chenyou Fan, Fangzheng Yan, Tian Li, Haitong Tang, et al. PRTS: A Primitive Reasoning and Tasking System via Contrastive Representations.arXiv preprint arXiv:2604.27472, 2026. URL https://arxiv.org/abs/2604.27472 on PithPith's reviewarXiv
  50. [70] Songming Liu, Bangguo Li, Kai Ma, Lingxuan Wu, Hengkai Tan, Xiao Ouyang, et al. RDT2: Exploring the Scaling Limit of UMI Data Towards Zero-Shot Cross-Embodiment Generalization.arXiv preprint arXiv:2602.03310, 2026. URL https://arxiv.org/abs/2602.03310 on PitharXiv
  51. [71] Delin Qu, Haoming Song, Qizhi Chen, Yuanqi Yao, Xinyi Ye, Yan Ding, et al. SpatialVLA: Exploring spatial representations for visual-language-action models. InRobotics: Science and Systems (RSS), 2025. doi: 10.15607/RSS.2025.XXI.011. URL https://www.roboticsproceedings.org/rss2... on Pithdoi:10.15607/rss.2025.xxi.011
  52. [72] StarVLA Community. StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing.arXiv preprint arXiv:2604.05014, 2026. URL https://arxiv.org/abs/2604.05014 on PithPith's reviewarXiv
  53. [73] Jun Cen, Siteng Huang, Yuqian Yuan, Kehan Li, Hangjie Yuan, Chaohui Yu, et al. RynnVLA-002: A Unified Vision-Language-Action and World Model.arXiv preprint arXiv:2511.17502, 2025. URL https://arxiv.org/abs/2511.17502 on PithPith's reviewarXiv
  54. [74] Tao Lin, Yilei Zhong, Yuxin Du, Jingjing Zhang, Jiting Liu, Yinxinyu Chen, et al. Evo-1: Lightweight vision-language-action model with preserved semantic alignment. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13397–13406, 2026. URL https://op... on Pith
  55. [77] Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, et al. SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics.arXiv preprint arXiv:2506.01844, 2025. URL https://arxiv.org/abs/2506.01844 on PithPith's reviewarXiv
  56. [78] Yi Yang, Zhihong Liu, Siqi Kou, Yiyang Chen, Yanzhe Hu, Jianbo Zhou, et al. World-Language-Action Model for Unified World Modeling, Language Reasoning, and Action Synthesis.arXiv preprint arXiv:2606.05979, 2026. URL https://arxiv.org/abs/2606.05979 on PithPith's reviewarXiv
  57. [79] Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, et al. X-VLA: Soft-prompted transformer as scalable cross-embodiment vision-language-action model. InInternational Conference on Learning Representations (ICLR), 2026. URL https://iclr.cc/virtual/... on Pith
  58. [80] URL https://arxiv.org/abs/2607.04988 on PithPith's reviewarXiv
  59. [81] Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, et al. Causal world modeling for robot control. InRobotics: Science and Systems (RSS), 2026. doi: 10.15607/RSS.2026.XXII.016. URL https://www.roboticsproceedings.org/rss22/p016.html on Pithdoi:10.15607/rss.2026.xxii.016
  60. [82] Sanjana Srivastava, Chengshu Li, Michael Lingelbach, Roberto Martín-Martín, Fei Xia, et al. BEHAVIOR: Benchmark for everyday household activities in virtual, interactive, and ecological environments. InConference on Robot Learning (CoRL), 2021 on Pith
  61. [83] Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, et al. RoboCasa: Large-scale simulation of household tasks for generalist robots. InRobotics: Science and Systems (RSS), 2024 on Pith
  62. [84] Haoquan Fang, Jiafei Duan, Donovan Clay, Sam Wang, Shuo Liu, Weikai Huang, et al. MolmoAct2: Action Reasoning Models for Real-world Deployment.arXiv preprint arXiv:2605.02881, 2026. URL https://arxiv.org/abs/2605.02881 on PithPith's reviewarXiv
  63. [88] Chao Yu, Yuanqing Wang, Zhen Guo, Hao Lin, Si Xu, Hongzhi Zang, Quanlu Zhang, Yongji Wu, Chunyang Zhu, Junhao Hu, Zixiao Huang, Mingjie Wei, Yuqing Xie, Ke Yang, Bo Dai, Zhexuan Xu, Jiakun Du, Xiangyuan Wang, Xu Fu, Letong Shi, Zhihao Liu, Kang Chen, Weilin Liu, Gang Liu, Boxu... on Pith
  64. [2023] doi: 10.1109/TRO.2023.3281153 on Pithdoi:10.1109/tro.2023.3281153
  65. [2024] doi: 10.1007/978-3-031-73383-3_21 on Pithdoi:10.1007/978-3-031-73383-3_21
  66. [2026] URL https://arxiv.org/abs/2602.11236 on PithPith's reviewarXiv

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