Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Zero2Skill claims that persistent language corrections let autonomous robot data collection need only 16% of teleoperation's human working time while matching its success rates.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 02:54 UTC pith:SCZF7EX6

load-bearing objection A genuinely fresh idea — persistent language corrections stored for the collection loop, not the policy — but the evidence is one small study, so treat the headline numbers as promising rather than proven. the 4 major comments →

arxiv 2607.14047 v3 pith:SCZF7EX6 submitted 2026-07-15 cs.RO cs.HCcs.SYeess.SY

Zero2Skill: Bootstrapping Robot Skills through Autonomous Data Collection, Training, and Deployment

Boyuan Wang , Zhenyuan Zhang , Zhiqin Yang , Peijun Gu , Shuya Wang , Xiaofeng Wang , Xianghui Ze , Yifan Chang
show 11 more authors
Guosheng Zhao Jiangnan Shao Guan Huang Hengyu Liu Yonggang Zhang Wei Xue Chunyuan Guan Chenglin Pu Yike Guo Xingang Wang Zheng Zhu
This is my paper
classification cs.RO cs.HCcs.SYeess.SY
keywords robot data collectionlanguage correctionscorrective memoryautonomous manipulationvision-language-action modelsinteractive autonomyverification-gated loophuman-robot symbiosis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Zero2Skill tries to show that the human cost of robot data collection can be made to scale with the number of distinct problems a session presents, not with the number of episodes or minutes. Its mechanism is a corrective memory: when the autonomous collect-verify-reset loop stalls, a remote operator sends one sentence, an LLM parses it into a structured rule, and that rule is reused on every later round. On a real desktop-clearing robot, this reduced human working time from 30.0 to 4.8 minutes for 50 demonstrations while keeping episode success at 100%, and a policy fine-tuned on the collected data reached the same 80% success as one trained on teleoperation data. A sympathetic reader would care because it points to a way of bootstrapping skills without continuous human teleoperation, using sparse language feedback as the control signal.

Core claim

The central claim is that a verification-gated collection loop combined with persistent language corrections can bootstrap a manipulation skill from a small number of human utterances. The system collects, verifies, and resets autonomously, pausing only when a phase exhausts a retry budget; the operator's free-form sentence is parsed into a structured adjustment stored in Corrective Memory, so a corrected failure mode typically does not need correction again under the same conditions. Zero2Skill reports matching teleoperation's episode collection success (100% vs 100%) and downstream policy success (80% vs 80%) while reducing human working time to 16% (4.8 vs 30.0 min), and language correcti

What carries the argument

The load-bearing object is Corrective Memory, a lightweight structured store of persistent corrections, each with a trigger, a correction, a scope, and the source utterance kept verbatim. An LLM parser decides whether an operator's sentence is a one-off adjustment or a reusable rule; reusable rules are consulted at the start of every round and applied when the trigger matches. The surrounding machinery is a verification-gated collect-reset loop with an explicit retry budget, plus offline re-judging of saved trajectories under corrected criteria, so language corrections propagate both to the collector's behavior and to the training set's labels.

Load-bearing premise

The load-bearing premise is that the LLM parser reliably turns free-form operator sentences into correctly conditioned, correctly scoped rules that stay valid in later rounds; if a stored rule is mis-conditioned, it can be reused wrongly, and the paper's own banana-and-chili run shows a correction that lowered success from 80% to 70.59% before partial recovery.

What would settle it

Run a long collection session with one recurring failure mode, apply a single operator correction, and count how many times the same failure re-triggers an alert in later rounds. If the alert rate does not drop to near zero after one stored correction, or if a regression test across 100 parser utterances shows a substantial share of rules stored with wrong triggers or scopes, the claim that oversight scales with distinct problems rather than session length would be contradicted.

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

If this is right

  • Human oversight during long-horizon collection can be reduced to interruptions for distinct problems rather than every recurrence, because corrections persist in memory.
  • Language correction of judging criteria improves VLM verifier accuracy in all four tested settings, with three reaching 10/10.
  • Execution-side language corrections improve single-attempt collection success, e.g., the approach-depth offset adds 22.5 points and arm-selection guidance adds 30.0 points, though stored corrections help only when they match the failure cause.
  • Policies fine-tuned on 50 Zero2Skill-collected trajectories match teleoperation-trained policy success (80% vs 80%), indicating autonomous data can be as useful for VLA fine-tuning.
  • Because each episode retains final observations, a corrected criterion can be applied retroactively, keeping the entire dataset labeled under one consistent standard.

Where Pith is reading between the lines

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

  • The 16% figure comes from a single task and a single collect-train-deploy cycle; whether oversight time continues to shrink across multiple cycles depends on the unvalidated multi-round flywheel claim.
  • If corrective memory rules transfer across tasks or embodiments, the per-task cost of human supervision could shrink further, but that rests on parser generalization beyond the tested settings.
  • The reported banana-and-chili regression (80% to 70.59% after the first correction) suggests a testable extension: add rollback or conflict-resolution when a stored rule lowers success, and measure whether cumulative performance improves.
  • The introduced Trajectories-per-Human-Minute axis could be applied to other collection pipelines to compare data utility per minute of human attention, although the paper only demonstrates it on one task.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. Zero2Skill is an autonomous robot data-collection system that combines a verification-gated collect–reset loop with a persistent Corrective Memory: when the loop fails repeatedly, a remote operator gives a natural-language correction, and an LLM parser converts the utterance into a structured rule that is reused on subsequent rounds. On a dual-arm desktop-clearing testbed the authors report that Zero2Skill collects 50 valid episodes with 100% episode success and 4.8 minutes of human working time (vs. 30.0 minutes for teleoperation), that language corrections improve verifier–human agreement in four settings and raise single-attempt collection success from 12.5% to 47.5%, and that a policy fine-tuned on Zero2Skill data matches a teleoperation-trained policy at 80% success. The paper's central claim is that persistent corrections make oversight cost grow with the number of distinct problems rather than session length, with a single collect–train–deploy cycle evaluated.

Significance. If the central claim holds, the paper makes a useful contribution: it shifts long-horizon data collection from episode-scoped human interventions to reusable, human-readable rules, and it provides an end-to-end evaluation of the collected data's utility for downstream policy learning. The experimental setup is real and the appendices define metrics, fine-tuning, and deployment protocols in unusual detail; the authors also honestly report a case where a stored correction fails to help (Table 3, banana-and-chili). However, the quantitative support is thin: the efficiency numbers rest on a single 50-episode run, the policy-equivalence claim rests on 20 blind trials, the verifier-accuracy table uses 10 judgment cases, and the LLM parser that underlies Corrective Memory is not evaluated. These gaps leave the headline efficiency and scaling claims plausible but not yet established.

major comments (4)
  1. [§4.3, Table 3] The banana-and-chili row is a direct counterexample to the central mechanism. After the first stored correction, cumulative success falls from 80.00% to 70.59% and never recovers above 76.47% despite two further interventions. The paper acknowledges "a stored correction is guaranteed to be reused, but not guaranteed to help," but it does not analyze why this correction misfired, whether the LLM parser mis-conditioned the trigger or scope, or how often such misfires occur. Without an error audit of Corrective Memory and an ablation with memory disabled, the claim that oversight scales with distinct problems rather than session length (Abstract, §1) is not established. This is load-bearing because the 16% human-time result depends on stored corrections being net beneficial.
  2. [§4.2, Table 1] The headline efficiency numbers come from a single 50-episode run per mode. The 4.8-minute human working time for Zero2Skill includes plan confirmation, language interventions, monitoring, and physical recovery, but no breakdown or variance is reported. With one run, the 16% ratio and TpHM=10.42 cannot be distinguished from run-to-run variation. Please report at least three independent sessions, or provide a per-component time budget (plan confirmation vs. language interventions vs. monitoring) and a sensitivity analysis to the retry limit N.
  3. [§4.4, Table 5] The claim that Zero2Skill data match teleoperation data rests on 80% vs 80% over 20 blind trials each. With n=20, the 95% confidence interval for an 80% estimate spans roughly 56–94%, so the observed point equality cannot support an equivalence claim. The same applies to the 55% scripted result. Please report confidence intervals, define an equivalence margin, or increase the number of trials; otherwise the data-quality claim is under-powered.
  4. [§3.2] The LLM parser is the sole channel through which the operator's linguistic intent becomes a persistent rule, yet the paper provides no evaluation of parser accuracy, no example of a complete parsed entry, and no analysis of parser-induced errors in the live collection runs. The Table 3 failure could be a parsing failure rather than a semantic mismatch. Please add a parser evaluation on held-out utterances (e.g., correct trigger/scope/correction extraction) and a qualitative audit of entries created during the reported sessions.
minor comments (5)
  1. [Abstract, §1] The text contains formatting artifacts with missing spaces (e.g., "WepresentZero2Skill,ahuman-robotsymbiotic..." and "Full Human Teleop Datasetting"), making it hard to read; please fix.
  2. [§4.1] The task description does not fully specify the object set, container geometry, placement distribution, or trial counts beyond "eight object categories"; Figure 4 lists objects but not precise configurations. Please add a full task specification for reproducibility.
  3. [§3.2] The four fields of Corrective Memory (trigger, correction, scope, source) are described but no concrete entry is shown. A single worked example (source utterance, parsed trigger/correction/scope) would make the mechanism concrete and help the reader assess parser behavior.
  4. [Appendix B.4] "Randomized initial placements" should state the randomization procedure and any seed, since the deployment trials form the basis of Table 5.
  5. [Table 3] The column "Object Exp. ID" is never defined; the four experiments should be tied to the experimental protocol. Also, because entries are cumulative, the marginal success rate in each intervention window should be reported alongside the cumulative rate.

Circularity Check

0 steps flagged

No significant circularity; results are empirical, externally evaluated, and do not reduce to their inputs.

full rationale

Zero2Skill is an empirical systems paper; its claims are measurements on a physical testbed rather than derivations from fitted parameters or self-referential equations. The 16% human-time figure (Table 1) is a direct measurement of operator attention against teleoperation and scripted baselines. The 80% vs 80% policy success (Table 5) comes from a blind deployment evaluation with the agent, verifier, and corrective memory removed at test time, so the result cannot be an artifact of the collection loop. The verifier-accuracy gains (Table 2) are the closest case to a concern: operators rewrite the judging criterion to match their own labels, and accuracy is measured against those same labels on overlapping cases. This is an in-sample calibration check rather than a prediction, and it is not forced by construction because the VLM (Seed1.8) must actually implement the revised prompt; the harder blue-box setting improves only to 4/10, showing the outcome is not tautological. The paper's own Table 3 discloses that a stored correction can reduce cumulative success (banana-and-chili: 80.00 -> 70.59), which is a limitation of the parser-dependent memory mechanism, not a circular step. The only self-citation (ClawNet, ref [28], whose authors overlap with the present paper) appears in the introduction as a general example of LLM agentic systems and is not load-bearing for any stated contribution; no uniqueness theorem or ansatz is imported from self-citations. External baselines provide independent comparison. Thus no load-bearing reduction of any claim to its own inputs is present.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The paper's central claims rest primarily on empirical evaluations rather than a derivation; no free parameters are fitted to the target success rates, but the retry limit and grasp-depth offset are hand-chosen. The main assumptions are about the reliability of the LLM/VLM components and the representativeness of the single-task testbed.

free parameters (2)
  • retry_limit_N = 3 (default)
    Set by hand prior to a run (Section 3.1, Algorithm 1). Determines the boundary between autonomous retry and human intervention; directly affects the human-time fraction and intervention frequency.
  • approach_depth_offset = not reported (operator-described 'one centimeter below')
    A persistent grasp-depth correction applied in the grasp tool (Appendix A.1). The +22.5 point single-attempt success gain in Figure 4 depends on this hand-chosen offset.
axioms (4)
  • domain assumption The VLM verifier (Seed1.8) returns reliable yes/no verdicts on fresh images after prompt correction.
    The entire verification-gated loop (Section 3.1) depends on the verifier's correctness; if the verifier is systematically wrong, bad trajectories are retained even with human correction.
  • domain assumption The LLM parser (DeepSeek-V3.2) maps free-form utterances into structured, correctly-conditioned adjustments with high fidelity.
    Section 3.2: the parser decides whether a correction persists and its scope. Table 3 shows a real case where a stored correction hurt, so this assumption is load-bearing and imperfect.
  • domain assumption Results on the single desktop-clearing task with a fixed object set generalize to broader manipulation tasks.
    All experiments use one task, one robot, and one object set (Section 4.1). The paper states limitations about complex dexterous tasks, so this assumption is not validated.
  • domain assumption Standard components (OpenClaw, SAM3, AnyGrasp, Pinocchio, ROS, cameras) function as documented.
    Appendix A describes the pipeline; if any component fails silently, the collection loop can degrade. The paper includes preflight checks to mitigate this.
invented entities (1)
  • Corrective Memory independent evidence
    purpose: Persistent store of structured language corrections consulted at the start of each collection round to modify verification prompts, grasp policies, object priorities, and retry limits.
    This is the core new system component. Its efficacy is tested in Experiments 2 and 3, with a negative result in Table 3, so it has direct falsifiable evidence within the paper.

pith-pipeline@v1.3.0-alltime-deepseek · 20476 in / 10822 out tokens · 99185 ms · 2026-08-02T02:54:33.812509+00:00 · methodology

0 comments
read the original abstract

Autonomous data collection governs the volume and quality of real-world trajectories for manipulation policy learning. Existing pipelines reduce human effort via self-resetting, VLM verification, or language-guided correction, yet episode-scoped fixes must be reissued whenever the same failure recurs, so oversight cost grows with session length rather than with the number of distinct problems. We present Zero2Skill, a human-robot symbiotic agentic system in which corrections are retained and reused across rounds. The collection loop collects, verifies, and resets autonomously, pausing for a remote operator only when a phase exhausts an explicit retry budget. An LLM parser maps each natural-language utterance to a structured adjustment stored in Corrective Memory, so addressed failure modes typically need not be corrected again under the same conditions. On a real-robot desktop-clearing testbed, Zero2Skill matches teleoperation episode success while reducing human working time to 16%. Language corrections improve verifier-human agreement in all four evaluated settings and raise average single-attempt success from 12.5% to 47.5% (arm-selection: 20.0% to 50.0%). Policies fine-tuned on Zero2Skill data match teleoperation-trained policy success at a fraction of collection human cost.

Figures

Figures reproduced from arXiv: 2607.14047 by Boyuan Wang, Chenglin Pu, Chunyuan Guan, Guan Huang, Guosheng Zhao, Hengyu Liu, Jiangnan Shao, Peijun Gu, Shuya Wang, Wei Xue, Xianghui Ze, Xiaofeng Wang, Xingang Wang, Yifan Chang, Yike Guo, Yonggang Zhang, Zheng Zhu, Zhenyuan Zhang, Zhiqin Yang.

Figure 1
Figure 1. Figure 1: (a) When each correction serves only the current episode, a recurring failure mode requires a new [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the agent-guided robotic data collection pipeline in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Timeline comparison of the three collection modes. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Single-attempt collection success rates across eight object categories under cumulative language [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representative observations and single-attempt collection success rates for language-guided arm [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Agent feedback for collection and reset verification. The system checks whether the grape bunch, [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Representative feedback for successful collection-reset episodes. The system checks whether the [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Representative scenes for the four verifier-correction settings evaluated in Table [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Per-episode distributions of Smoothness, Aesthetics, Brightness, and RMS jerk for the agent-collected [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 16 linked inside Pith

  1. [1]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Lucy Xiaoyang Shi, James Tanner, Quan Vuong, Anna Walling, Haohuan Wang, and Ury Zhilinsky.𝜋0: A visio...

    Pith/arXiv arXiv 2024

  2. [2]

    Scaling robot policy learning via zero-shot labeling with foundation models.arXiv preprint arXiv:2410.17772, 2024

    Nils Blank, Moritz Reuss, Marcel Rühle, Ömer Erdinç Yağmurlu, Fabian Wenzel, Oier Mees, and Rudolf Lioutikov. Scaling robot policy learning via zero-shot labeling with foundation models.arXiv preprint arXiv:2410.17772, 2024. 6

    Pith/arXiv arXiv 2024

  3. [3]

    Seed1.8 model card: Towards generalized real-world agency, 2026

    Bytedance Seed. Seed1.8 model card: Towards generalized real-world agency, 2026. URLhttps: //arxiv.org/abs/2603.20633. 9

    Pith/arXiv arXiv 2026

  4. [4]

    Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025. 11, 19

    Pith/arXiv arXiv 2025

  5. [5]

    The pinocchio c++ library – a fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives

    Justin Carpentier, Guilhem Saurel, Gabriele Buondonno, Joseph Mirabel, Florent Lamiraux, Olivier Stasse, and Nicolas Mansard. The pinocchio c++ library – a fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives. InIEEE International Symposium on System Integrations (SII), 2019. 19

    2019

  6. [6]

    Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots

    Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. InProceedings of Robotics: Science and Systems (RSS), 2024. 2, 3

    2024

  7. [7]

    DeepSeek-V3.2: Pushing the frontier of open large language models, 2025

    DeepSeek-AI. DeepSeek-V3.2: Pushing the frontier of open large language models, 2025. URLhttps: //arxiv.org/abs/2512.02556. 9

    Pith/arXiv arXiv 2025

  8. [8]

    Anygrasp: Robust and efficient grasp perception in spatial and temporal domains

    Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 39(5):3929–3945, 2023. 19

    2023

  9. [9]

    Brown, and Ken Goldberg

    Ryan Hoque, Ashwin Balakrishna, Ellen Novoseller, Albert Wilcox, Daniel S. Brown, and Ken Goldberg. ThriftyDAgger: Budget-aware novelty and risk gating for interactive imitation learning. In5th Conference on Robot Learning (CoRL), volume 164 ofProceedings of Machine Learning Research, pages 598–608. PMLR, 2022a. 2, 4

  10. [10]

    Fleet-DAgger: Interactive robot fleet learning with scalable human supervision

    Ryan Hoque, Lawrence Yunliang Chen, Satvik Sharma, Karthik Dharmarajan, Brijen Thananjeyan, Pieter Abbeel, and Ken Goldberg. Fleet-DAgger: Interactive robot fleet learning with scalable human supervision. In6th Conference on Robot Learning (CoRL), volume 205 ofProceedings of Machine Learning Research. PMLR, 2022b. 2, 4

  11. [11]

    ABot-Claw: A foundation for persistent, cooperative, and self-evolving robotic agents, 2026

    Dongjie Huo, Haoyun Liu, Guoqing Liu, Dekang Qi, Zhiming Sun, Maoguo Gao, Jianxin He, Yandan Yang, Xinyuan Chang, Feng Xiong, Xing Wei, Zhiheng Ma, and Mu Xu. ABot-Claw: A foundation for persistent, cooperative, and self-evolving robotic agents, 2026. URLhttps://arxiv.org/abs/2604.10096. 4

    Pith/arXiv arXiv 2026

  12. [12]

    DROID: A large- scale in-the-wild robot manipulation dataset

    Alexander Khazatsky, Karl Pertsch, Suraj Nair, Ashwin Balakrishna, Sudeep Dasari, Siddharth Karamcheti, Soroush Nasiriany, Mohan Kumar Srirama, Lawrence Yunliang Chen, Kirsty Ellis, et al. DROID: A large- scale in-the-wild robot manipulation dataset. InProceedings of Robotics: Science and Systems (RSS), 2024. 1, 3 16 Zero2Skill: Bootstrapping Robot Skills...

    2024

  13. [13]

    OpenVLA: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. OpenVLA: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2...

    Pith/arXiv arXiv 2024

  14. [14]

    IROSA: Interactive robot skill adaptation using natural language.IEEE Robotics and Automation Letters,

    Markus Knauer, Samuel Bustamante, Thomas Eiband, Alin Albu-Schäffer, Freek Stulp, and João Silvério. IROSA: Interactive robot skill adaptation using natural language.IEEE Robotics and Automation Letters,

  15. [15]

    RoboClaw: An agentic framework for scalable long-horizon robotic tasks

    Ruiying Li, Yunlang Zhou, YuYao Zhu, Kylin Chen, Jingyuan Wang, Sukai Wang, Kongtao Hu, Minhui Yu, Bowen Jiang, Zhan Su, Jiayao Ma, Xin He, Yongjian Shen, Yang Yang, Guanghui Ren, Maoqing Yao, Wenhao Wang, and Yao Mu. RoboClaw: An agentic framework for scalable long-horizon robotic tasks. arXiv preprint arXiv:2603.11558, 2026. 2, 3, 6

    arXiv 2026

  16. [16]

    Multi-task interactive robot fleet learning with visual world models

    Huihan Liu, Yu Zhang, Vaarij Betala, Evan Zhang, James Liu, Crystal Ding, and Yuke Zhu. Multi-task interactive robot fleet learning with visual world models. In8th Conference on Robot Learning (CoRL),

  17. [17]

    Towards end-to-end automation of ai research.Nature, 651(8107):914–919, 2026a

    Chris Lu, Cong Lu, Robert Tjarko Lange, Yutaro Yamada, Shengran Hu, Jakob Foerster, David Ha, and Jeff Clune. Towards end-to-end automation of ai research.Nature, 651(8107):914–919, 2026a. 2

  18. [18]

    ASPIRE: Agentic skills discovery for robotics, 2026b

    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, 2026b. URLhttps://arxiv.org/abs/2607.00272. 4

    Pith/arXiv arXiv

  19. [19]

    Skillclaw: Let skills evolve collectively with agentic evolver.arXiv preprint arXiv:2604.08377,

    Ziyu Ma, Shidong Yang, Yuxiang Ji, Xucong Wang, Yong Wang, Yiming Hu, Tongwen Huang, and Xiangx- iang Chu. Skillclaw: Let skills evolve collectively with agentic evolver.arXiv preprint arXiv:2604.08377,

    Pith/arXiv arXiv

  20. [20]

    So you think you can scale up autonomous robot data collection? In8th Conference on Robot Learning (CoRL),

    Suvir Mirchandani, Suneel Belkhale, Joey Hejna, Evelyn Choi, Md Sazzad Islam, and Dorsa Sadigh. So you think you can scale up autonomous robot data collection? In8th Conference on Robot Learning (CoRL),

  21. [21]

    Open X-Embodiment: Robotic learning datasets and RT-X models.arXiv preprint arXiv:2310.08864,

    Open X-Embodiment Collaboration, Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, Antonin Raffin, et al. Open X-Embodiment: Robotic learning datasets and RT-X models.arXiv preprint arXiv:2310.08864,

    Pith/arXiv arXiv

  22. [22]

    𝜋0.5: a vision-language-action model with open-world generalization, 2025

    Physical Intelligence. 𝜋0.5: a vision-language-action model with open-world generalization, 2025. URL https://arxiv.org/abs/2504.16054. 8, 14, 22

    Pith/arXiv arXiv 2025

  23. [23]

    Zhao, Archit Sharma, Karl Pertsch, Jianlan Luo, Sergey Levine, and Chelsea Finn

    Lucy Xiaoyang Shi, Zheyuan Hu, Tony Z. Zhao, Archit Sharma, Karl Pertsch, Jianlan Luo, Sergey Levine, and Chelsea Finn. Yell at your robot: Improving on-the-fly from language corrections.arXiv preprint arXiv:2403.12910, 2024. 2, 4

    Pith/arXiv arXiv 2024

  24. [24]

    Openclaw: Your own personal ai assistant, 2024

    Peter Steinberger and contributors. Openclaw: Your own personal ai assistant, 2024. URLhttps: //github.com/openclaw/openclaw. Available at https://openclaw.ai. 2, 4

    2024

  25. [25]

    RADAR: Closed- loop robotic data generation via semantic planning and autonomous causal environment reset.arXiv preprint arXiv:2603.11811, 2026

    Yongzhong Wang, Keyu Zhu, Yong Zhong, Liqiong Wang, Jinyu Yang, and Feng Zheng. RADAR: Closed- loop robotic data generation via semantic planning and autonomous causal environment reset.arXiv preprint arXiv:2603.11811, 2026. 2, 4, 6

    Pith/arXiv arXiv 2026

  26. [26]

    Max" Fu, Haoru Xue, Jalen Lu, Yi Yang, Cunxi Dai, Zi Wang, Jimmy Wu, Guanzhi Wang, S. Shankar Sastry, Ken Goldberg, Linxi

    Wenli Xiao, Jia Xie, Tonghe Zhang, Haotian Lin, Letian "Max" Fu, Haoru Xue, Jalen Lu, Yi Yang, Cunxi Dai, Zi Wang, Jimmy Wu, Guanzhi Wang, S. Shankar Sastry, Ken Goldberg, Linxi "Jim" Fan, Yuke Zhu, and Guanya Shi. ENPIRE: Agentic robot policy self-improvement in the real world, 2026. URL https://arxiv.org/abs/2606.19980. 4 17 Zero2Skill: Bootstrapping Ro...

    Pith/arXiv arXiv 2026

  27. [27]

    A-mem: Agentic memory for llm agents.Advances in Neural Information Processing Systems, 38:17577–17604, 2026

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-mem: Agentic memory for llm agents.Advances in Neural Information Processing Systems, 38:17577–17604, 2026. 2

    2026

  28. [28]

    ClawNet: Human-symbiotic agent network for cross-user autonomous cooperation.arXiv preprint arXiv:2604.19211, 2026

    Zhiqin Yang, Zhenyuan Zhang, Xianzhang Jia, Jun Song, Wei Xue, Yonggang Zhang, and Yike Guo. ClawNet: Human-symbiotic agent network for cross-user autonomous cooperation.arXiv preprint arXiv:2604.19211, 2026. 2

    Pith/arXiv arXiv 2026

  29. [29]

    Distilling and retrieving generalizable knowledge for robot manipulation via language corrections

    Lihan Zha, Yuchen Cui, Li-Heng Lin, Minae Kwon, Montserrat Gonzalez Arenas, Andy Zeng, Fei Xia, and Dorsa Sadigh. Distilling and retrieving generalizable knowledge for robot manipulation via language corrections. InIEEE International Conference on Robotics and Automation (ICRA), 2024. 2, 4

    2024

  30. [30]

    A survey on the memory mechanism of large language model-based agents.ACM Transactions on Information Systems, 43(6):1–47, 2025

    Zeyu Zhang, Quanyu Dai, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. A survey on the memory mechanism of large language model-based agents.ACM Transactions on Information Systems, 43(6):1–47, 2025. 2

    2025

  31. [31]

    Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn

    Tony Z. Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual ma- nipulation with low-cost hardware. InProceedings of Robotics: Science and Systems (RSS), 2023. 2, 3

    2023

  32. [32]

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

    Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag Sanketi, Gracia Salazar, Michael S Ryoo, et al. RT-2: Vision- language-action models transfer web knowledge to robotic con...

    2023

  33. [2026]

    arXiv preprint arXiv:2603.03897. 2, 4

    Pith/arXiv arXiv