Pith. sign in

REVIEW 2 major objections 5 minor 300 references

The paper claims robot learning is splitting into frozen-weight VLA policies versus self-improving code-as-policy agents, and that only a handful of recent systems combine execution feedback, persistent skill memory, and evolutionary search

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

2026-08-04 19:37 UTC pith:ZVOL4XMT

load-bearing objection Useful new framing, but the headline sparsity claim overstates what a curated corpus can support. the 2 major comments →

arxiv 2608.01851 v1 pith:ZVOL4XMT submitted 2026-08-03 cs.RO cs.AI

Weights or Skills? A Survey of Robot-Learning Techniques: from Action-Predicting Weights to Robots that Write their Own Skills

Gaytri Jena , Kapil Wanaskar , Vinija Jain , Aman Chadha , Vasu Sharma , Amitava Das This is my paper
classification cs.RO cs.AI
keywords robot learningcode-as-policyself-improvementvision-language-action modelsskill librariesskill economysurvey
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.

This survey argues that the field of robot learning is organising itself around one question: should competence ship as frozen network weights (vision-language-action models) or as executable skills (code-as-policy agents)? The paper's central analytical move is to arrange code-as-policy methods on a five-rung ladder of self-improvement, from one-shot program synthesis up to a full loop that combines execution feedback, persistent skill memory, and evolutionary search. That top rung is the least populated cell in the field, occupied by only a few recent systems such as ASPIRE, ENPIRE, and RoboClaw. The survey also disentangles five senses of the word 'skill', shows that only the code sense self-improves without gradient updates, and connects the taxonomy to commercial robot-skill marketplaces, which today ship only static playback. A sympathetic reader would care because the paper names where the field's open frontier is and gives a concrete measurement agenda for reaching it.

Core claim

Degree of self-improvement, not application or embodiment, is the right axis for code-as-policy robot learning. Systems that write control code fall into five rungs: zero-shot synthesis, closed-loop self-repair, skill-library accumulation, evolutionary program search, and the full loop where feedback, memory, and search operate together. The paper's structural observation is that the top rung is nearly empty; only three recent systems satisfy all three mechanisms (F, M, and S). Frozen-weight VLA policies have no code and no loop, so their competence is fixed at training and cannot be edited, audited, or recombined. Of the five senses of 'skill', only the code sense is simultaneously inspecta

What carries the argument

The self-improvement ladder and its three tagged mechanisms: Feedback (an execution-grounded signal revises the code within a task), Memory (validated code persists across task boundaries), and Search (multiple candidate programs are maintained, scored, and selected or mutated). Table 3 gives necessary-and-sufficient conditions for each tag, and the ladder's fifth rung is defined as the conjunction F, M, and S, the full self-improving loop. The machinery does two jobs: it classifies every code-as-policy system into an unambiguous rung, and it makes the sparsity of the top cell a checkable claim rather than an impression.

Load-bearing premise

The load-bearing premise is that the curated 77-system corpus plus the qualitative placement rules are complete enough that the emptiness of the top cell reflects the field, not the selection.

What would settle it

Find one code-as-policy system outside the survey's 77 that, on the paper's own definitions, reads an execution-grounded signal to revise its program, stores validated code across task boundaries, and maintains a scored population of candidate programs; its existence would put a fourth system in the sparsely populated top cell and falsify the sparsity claim.

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

If this is right

  • If the ladder is the right axis, code-as-policy systems are comparable by how much competence they produce at run time, and the field's frontier is the feedback-plus-memory-plus-search cell.
  • Only three systems currently occupy that cell, making it the least surveyed and least populated region of the map.
  • Because only code-form skills are inspectable, adaptable, and composable at once, while commercial app-form skills are static, closing the marketplace gap requires applying the full loop to downloaded skills.
  • No standard benchmark yet measures whether a robot improves with experience; the proposed success-versus-interactions curve and skill-library reuse rate would turn 'self-improvement' from a claim into a reported number.
  • Frozen-weight VLA policies have no code and no loop, so they cannot be edited, audited, or recombined after training, which is exactly what the skills pole offers.

Where Pith is reading between the lines

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

  • The emptiness of the top rung may reflect the survey's corpus-construction rules as much as the field; a wider sweep could find additional F+M+S systems, so the sparsity claim is best read as a testable hypothesis.
  • If full-loop self-improvement becomes dependable, skill marketplaces would shift from selling fixed playback to selling a base skill plus an on-device adaptation loop, making skills assets that appreciate with use.
  • The five-sense skill analysis points to a design target the survey does not itself build: a skill representation that is simultaneously inspectable code and a distributable, versioned package with declared pre-conditions and effects.
  • A concrete next step the survey calls for but does not run: apply its four-metric protocol (success vs. interactions, reuse rate, transfer drop, provenance coverage) to ASPIRE, ENPIRE, and RoboClaw under a common interaction budget.

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

2 major / 5 minor

Summary. This survey organizes robot learning around a 'weights vs. skills' axis. Using PRISMA-style selection, it builds a 77-system taxonomy (plus a 225-work landscape appendix) across six families, and its central analytical contribution is a code-as-policy 'self-improvement ladder' defined by three mechanisms — Feedback, Memory, and Search (Table 3) — with rungs from zero-shot program synthesis through closed-loop self-repair, skill-library accumulation, and evolutionary search to the full F+M+S loop. The headline claim is that this top cell is sparsely populated, occupied only by three concurrent systems, ASPIRE, ENPIRE, and RoboClaw. The paper also distinguishes five senses of 'skill' (Table 5), positions reward synthesis, sim-to-real, and benchmark families relative to the two poles, and links the taxonomy to commercial robot-skill marketplaces, closing with a measurable evaluation protocol (Table 9).

Significance. Strengths worth naming: the operational F/M/S definitions, with an explicit 'does not count' column (Table 3), are a reusable instrument and make the sparsity claim falsifiable in principle; the five-senses analysis of 'skill' (Table 5) sharpens a genuinely overloaded term; Table 9 converts self-improvement into reportable quantities; and the survey is unusually transparent — placement criteria (§2.1), honest logging of what was not logged (Figure 4), and per-panel figure provenance (Table 10). If the taxonomy holds, it gives the community a shared vocabulary for a fast-moving area. The caveat is that the headline sparsity claim currently asserts more than the curated corpus can verify; the major comments address that gap.

major comments (2)
  1. [Abstract; §3.1.5; Conclusion; §2.1; Figure 4] The sparsity of the F+M+S cell is stated as a field-level structural observation ('the least populated ... region of the field', Abstract and §7), but the evidence is corpus-level. §2.1's placement rules taxonomize only 'distinct' exemplars; near-variants and reimplementations are relegated to the landscape appendix without F/M/S scoring. Figure 4's right (taxonomy) column reports no candidate counts or exclusions, and metadata-failed candidates were 'discarded at harvest time; their count was not logged.' A reader therefore cannot rule out F+M+S systems outside the 77. The claim should be qualified to 'within the surveyed corpus' wherever it appears, or backed by an explicit sweep of the landscape corpus against Table 3 with excluded candidates logged. The three occupants are concurrent 2026 preprints, so the field-level framing should carry the §5 snapshot hedge.
  2. [Table 4; §3.1.3; §3.1.5; Table 3] The F/M/S annotations in Table 4 are asserted without justification, and at least one conflicts with Table 3: Voyager is marked F=✘ although its published mechanism ('iterative prompting with environment feedback') revises code from execution-grounded signals within a task, which Table 3 counts as Feedback. Since Voyager is the paper's canonical memory-rung example, the consistent application of the definitions is in question. The top-rung rows need particular support: §3.1.5 describes RoboClaw only as 'self-resetting loops', which does not obviously satisfy the Table 3 Search condition ('more than one candidate program ... scored by execution ... selected or mutated'). For the cell whose sparsity is the paper's headline, per-system evidence from the primary sources — or an explicit interpretive note — is required.
minor comments (5)
  1. [§3.1; Table 4] The 'rungs' are not cumulative: rung 3 (M present, F optional) and rung 4 (F+S, M absent) are incomparable on the paper's own axes. The five 'yes' questions in §3.1 read as a sequence, but Table 4's structure is a path through the F/M/S power set. Please clarify what 'degree of self-improvement' orders, or state that the rungs are conjunctions with a narrative ordering.
  2. [Table 4, §3.1.3 block header] The block header row 'F ✘ M ✔ S ✘' contradicts the DROC row (F ✔) within the same block. Relabel the block as 'M present; F optional; S absent' to match the membership condition in §3.1.3.
  3. [§3.3] 'What ships is still weights' is imprecise for Eureka, which ships reward code; the separation from §3.1 rests on the ladder's first question (whether the code directly produces control actions). Please state this boundary explicitly.
  4. [Figure 4] The taxonomy (right) column reports no candidate or exclusion counts. For parity with the landscape stream's transparency, either log these numbers or state explicitly that they were not logged.
  5. [§3.6; Table 11] Typography: 'BEHA VIOR-1K' (§3.6), 'V ALOR[1]', and 'A W AC[188]' (Table 11) contain stray spaces.

Circularity Check

0 steps flagged

No circularity: the taxonomy is author-constructed and corpus-scoped, but no prediction reduces to its inputs and no load-bearing self-citation appears.

full rationale

The paper is a survey whose organizing claim is a taxonomy: code-as-policy methods are arranged by degree of self-improvement, and the top feedback+memory+search cell is said to be sparsely populated. This is not a derivation or prediction; it is an explicitly framed analytical classification. The authors state the corpus is deliberately focused ('This is a deliberately focused survey. Rather than cataloguing the field exhaustively, it examines 77 representative systems...', Abstract) and that the 77/225 split is about analytical role ('The 77/225 split is therefore a statement about analytical role, sharp exemplar versus representative breadth, not about a work's quality or importance', §2.1). The F/M/S operational definitions in Table 3 are author-supplied criteria, and Table 4's caption notes the mechanism columns are 'determined by the rung'—a presentational tautology, not evidence that the rung assignments were derived from those columns. The actual placements are argued from external system descriptions (§3.1.5 describes ASPIRE, ENPIRE, and RoboClaw as combining execution traces, skill libraries, and population search). The field-level wording in the abstract and conclusion is broader than the curated-corpus evidence, but §5 explicitly limits the survey to a 'snapshot' and acknowledges the taxonomy is code-centric. No fitted parameter is renamed as a prediction, no equation-level reduction exists, and the reference list contains no works by the present authors, so there is no load-bearing self-citation or imported uniqueness theorem. The main caveat—that the sparsity claim is scoped to a curated, non-exhaustive corpus—is a correctness/coverage concern, not circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 0 invented entities

No free parameters are fitted because this is a survey, not a model. The analytical machinery consists of author-defined categories (weights vs skills, F/M/S, five senses of skill) that are not derived from equations or data. The principal load-bearing assumptions are the representativeness of the curated corpus and the correctness of the qualitative placement of systems into the taxonomy.

axioms (4)
  • domain assumption The 77 taxonomy systems are representative enough to support claims about the relative emptiness of the top self-improvement rung.
    The corpus is curated by seeding and snowballing (§2.1), not by exhaustive search; the sparsity claim generalizes only if the curation missed no F+M+S systems.
  • ad hoc to paper The operational definitions of Feedback, Memory, and Search (Table 3) carve the field at its natural joints.
    These definitions are constructed by the authors for this survey; there is no independent benchmark showing they are the correct decomposition of self-improvement.
  • domain assumption The cited systems, including the 2026 preprints ASPIRE, ENPIRE, and RoboClaw, behave as the survey reports.
    Metadata verification is claimed (§2.1), but no code, data, or commit hashes are shipped, so a reader cannot independently verify the F/M/S attributes.
  • domain assumption A single commercial marketplace (Unitree UniStore) is a sufficient basis for describing an 'emerging skill economy'.
    §4 builds the open-problems agenda on this example; the generality of the trend is assumed.

pith-pipeline@v1.3.0-daily-deepseek · 43458 in / 11705 out tokens · 129705 ms · 2026-08-04T19:37:46.854077+00:00 · methodology

0 comments
read the original abstract

Robot learning is splitting into two bets: policies that bake competence into frozen weights (vision-language-action, or VLA, models), and agents that write and refine their own executable skills as code. This survey organises the field around that axis of weights versus skills. Its central analytical contribution is a deep-dive that arranges code-as-policy methods by their degree of self-improvement, from zero-shot program synthesis, through closed-loop self-repair and persistent skill memory, to the sparsely populated cell in which execution feedback, skill memory, and evolutionary search combine into one open-ended loop; only a few very recent systems (for example ASPIRE, ENPIRE, and RoboClaw) occupy that cell. We map the complementary "skills" pole, from unsupervised reinforcement-learning skill discovery to large-language-model skill libraries, and show that the word "skill" is used in at least five distinct senses, of which only the code sense self-improves without gradient updates. We then connect the taxonomy to the emerging skill economy: commercial robot-skill marketplaces now distribute one-tap skills across robots but ship only static playback, which surfaces open problems of adaptation, cross-embodiment portability, provenance, safety verification, composition, and standardisation. This is a deliberately focused survey. Rather than cataloguing the field exhaustively, it examines 77 representative systems across six technique families through one taxonomy and a set of contrast tables, and it supplies operational definitions of the self-improvement mechanisms together with a statement of what each family cannot do.

Figures

Figures reproduced from arXiv: 2608.01851 by Aman Chadha, Amitava Das, Gaytri Jena, Kapil Wanaskar, Vasu Sharma, Vinija Jain.

Figure 1
Figure 1. Figure 1: Weights or Skills? The survey’s organising question, posed with real systems. Weights (left): vision￾language-action policies ship frozen network weights that map observations to actions (§3.2). Skills (right): code-as-policy agents ship executable programs and improve them from their own experience (§3.1). The 77 taxonomy systems populate these two poles across six technique branches (2016–2026). Every ph… view at source ↗
Figure 2
Figure 2. Figure 2: The weights-versus-skills taxonomy of robot learning (all 77 systems). Each sub-family is one cell listing its systems, each with a trimmed form of the paper’s own title. Left fork = inspectable code / skills (§3.1, §3.4); the §3.1 cells are shaded by degree of self-improvement, from zero-shot to the full feedback+memory+search loop . Full per-system detail: Tables 2–6 [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 3
Figure 3. Figure 3: The closest related surveys as a timeline (2024–2026); [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Corpus construction in the PRISMA 2020 style [196]. The left column is the web-search stream behind the landscape corpus ( [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The surveyed corpus at a glance. The survey covers 302 systems (77 placed in the taxonomy of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Profile of the 225 landscape works ( [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The two poles over time. Year of first release of the taxonomy systems this survey places at each pole of [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Robots across the surveyed systems. Representative embodiments from the corpus, grouped into real-world manipulators, mobile/legged/aerial platforms, and simulation embodiments. Systems from both poles of the weights–skills axis ( [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: p.9 Voyager Fig. A.4 p.42 DIAYN [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The full self-improving loop (§3.1.5): feedback (F) [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: An actionable evaluation protocol for the future agenda (§6; [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 8
Figure 8. Figure 8: robots (27 panels) RT-1 arXiv:2212.06817 [PITH_FULL_IMAGE:figures/full_fig_p040_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: results (20 panels) RT-2 arXiv:2307.15818 Fig. 6a rt2_emergent_dm.png Code-as-Policies arXiv:2209.07753 [PITH_FULL_IMAGE:figures/full_fig_p040_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

300 extracted references · 1 canonical work pages

  1. [1]

    Joshua Achiam, Harrison Edwards, Dario Amodei, and Pieter Abbeel. 2018. Variational Option Discovery Algorithms. arXiv preprint arXiv:1807.10299 (2018)

    Pith/arXiv arXiv 2018

  2. [2]

    Ademi Adeniji, Amber Xie, Carmelo Sferrazza, Younggyo Seo, Stephen James, and Pieter Abbeel. 2023. Language Reward Modulation for Pretraining Reinforcement Learning. arXiv preprint arXiv:2308.12270 (2023). Weights or Skills? 1:21 Metric Definition Instantiation (testbed) Family stressed Success-vs- interactions curve held-out success rate as a function of...

    Pith/arXiv arXiv 2023

  3. [3]

    Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. 2022. Legged Locomotion in Challenging Terrains using Egocentric Vision. In Conference on Robot Learning (CoRL) . arXiv:2211.07638

    Pith/arXiv arXiv 2022

  4. [4]

    Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, et al. 2022. Do As I Can, Not As I Say: Grounding Language in Robotic Affordances. arXiv preprint arXiv:2204.01691 (2022)

    Pith/arXiv arXiv 2022

  5. [5]

    Anurag Ajay et al. 2023. Is Conditional Generative Modeling all you need for Decision-Making?. In International Conference on Learning Representations (ICLR) . arXiv:2211.15657

    Pith/arXiv arXiv 2023

  6. [6]

    Anurag Ajay, Aviral Kumar, Pulkit Agrawal, Sergey Levine, and Ofir Nachum. 2021. OPAL: Offline Primitive Discovery for Accelerating Offline Reinforcement Learning. In International Conference on Learning Representations (ICLR) . arXiv:2010.13611

    Pith/arXiv arXiv 2021

  7. [7]

    Ilge Akkaya et al. 2019. Solving Rubik’s Cube with a Robot Hand. arXiv preprint arXiv:1910.07113 (2019)

    Pith/arXiv arXiv 2019

  8. [8]

    Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce, and François Fleuret. 2024. Diffusion for World Modeling: Visual Details Matter in Atari. In Advances in Neural Information Processing Systems (NeurIPS). arXiv:2405.12399

    Pith/arXiv arXiv 2024

  9. [9]

    Marcin Andrychowicz et al. 2018. Learning Dexterous In-Hand Manipulation. arXiv preprint arXiv:1808.00177 (2018)

    Pith/arXiv arXiv 2018

  10. [10]

    Pierre-Luc Bacon, Jean Harb, and Doina Precup. 2017. The Option-Critic Architecture. In Proceedings of the AAAI Conference on Artificial Intelligence. arXiv:1609.05140

    Pith/arXiv arXiv 2017

  11. [11]

    Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. 2023. Affordances from Human Videos as a Versatile Representation for Robotics. InIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . arXiv:2304.08488

    Pith/arXiv arXiv 2023

  12. [12]

    Homanga Bharadhwaj et al. 2024. RoboAgent: Generalization and Efficiency in Robot Manipulation via Seman- tic Augmentations and Action Chunking. In IEEE International Conference on Robotics and Automation (ICRA) . arXiv:2309.01918

    Pith/arXiv arXiv 2024

  13. [13]

    Kevin Black et al . 2024. Zero-Shot Robotic Manipulation with Pretrained Image-Editing Diffusion Models. In International Conference on Learning Representations (ICLR) . arXiv:2310.10639

    Pith/arXiv arXiv 2024

  14. [14]

    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 LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Allen Z. Ren, Lucy Xi...

    Pith/arXiv arXiv 2025

  15. [15]

    and Ury Zhilinsky

    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, 1:22 Jena et al. and Ury Zhil...

    Pith/arXiv arXiv 2024

  16. [16]

    Lee, Maria Bauza, et al

    Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Devin, Alex X. Lee, Maria Bauza, et al. 2023. RoboCat: A Self-Improving Generalist Agent for Robotic Manipulation. Transactions on Machine Learning Research (TMLR) (2023). arXiv:2306.11706

    Pith/arXiv arXiv 2023

  17. [17]

    Anthony Brohan, Noah Brown, Justice Carbajal, et al. 2023. RT-1: Robotics Transformer for Real-World Control at Scale. In Robotics: Science and Systems (RSS) . arXiv:2212.06817

    Pith/arXiv arXiv 2023

  18. [18]

    Anthony Brohan, Noah Brown, Justice Carbajal, et al. 2023. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. In Conference on Robot Learning (CoRL) . arXiv:2307.15818

    Pith/arXiv arXiv 2023

  19. [19]

    Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, et al. 2024. Genie: Generative Interactive Environ- ments. In International Conference on Machine Learning (ICML) . arXiv:2402.15391

    Pith/arXiv arXiv 2024

  20. [20]

    Victor Campos, Alexander Trott, Caiming Xiong, Richard Socher, Xavier Giro-i Nieto, and Jordi Torres. 2020. Explore, Discover and Learn: Unsupervised Discovery of State-Covering Skills. InInternational Conference on Machine Learning (ICML). arXiv:2002.03647

    Pith/arXiv arXiv 2020

  21. [21]

    Thomas Carta, Clément Romac, Thomas Wolf, Sylvain Lamprier, Olivier Sigaud, and Pierre-Yves Oudeyer. 2023. Grounding Large Language Models in Interactive Environments with Online Reinforcement Learning. arXiv preprint arXiv:2302.02662 (2023)

    arXiv 2023

  22. [22]

    Chi-Lam Cheang et al. 2024. GR-2: A Generative Video-Language-Action Model with Web-Scale Knowledge for Robot Manipulation. arXiv preprint arXiv:2410.06158 (2024)

    Pith/arXiv arXiv 2024

  23. [23]

    Yevgen Chebotar et al. 2023. Q-Transformer: Scalable Offline Reinforcement Learning via Autoregressive Q-Functions. In Conference on Robot Learning (CoRL) . arXiv:2309.10150

    Pith/arXiv arXiv 2023

  24. [24]

    Yevgen Chebotar, Ankur Handa, Viktor Makoviychuk, Miles Macklin, Jan Issac, Nathan Ratliff, and Dieter Fox. 2019. Closing the Sim-to-Real Loop: Adapting Simulation Randomization with Real World Experience. InIEEE International Conference on Robotics and Automation (ICRA) . arXiv:1810.05687

    Pith/arXiv arXiv 2019

  25. [25]

    Chang Chen, Yi-Fu Wu, Jaesik Yoon, and Sungjin Ahn. 2022. TransDreamer: Reinforcement Learning with Transformer World Models. arXiv preprint arXiv:2202.09481 (2022)

    Pith/arXiv arXiv 2022

  26. [26]

    Harold Haodong Chen, Sirui Chen, Yingjie Xu, Wenhang Ge, and Ying-Cong Chen. 2026. RoboEvolve: Co-Evolving Planner-Simulator for Robotic Manipulation with Limited Data. arXiv preprint (2026). arXiv:2605.13775

    Pith/arXiv arXiv 2026

  27. [27]

    Junting Chen, Yao Mu, Qiaojun Yu, Tianming Wei, Silang Wu, Zhecheng Yuan, Zhixuan Liang, Chao Yang, Kaipeng Zhang, Wenqi Shao, Yu Qiao, Huazhe Xu, Mingyu Ding, and Ping Luo. 2024. RoboScript: Code Generation for Free-Form Manipulation Tasks across Real and Simulation. arXiv preprint (2024). arXiv:2402.14623

    Pith/arXiv arXiv 2024

  28. [28]

    Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. 2021. Decision Transformer: Reinforcement Learning via Sequence Modeling. In Advances in Neural Information Processing Systems (NeurIPS) . arXiv:2106.01345

    Pith/arXiv arXiv 2021

  29. [29]

    Lawrence Yunliang Chen, Kush Hari, Karthik Dharmarajan, Chenfeng Xu, Quan Vuong, and Ken Goldberg. 2024. Mirage: Cross-Embodiment Zero-Shot Policy Transfer with Cross-Painting. In Robotics: Science and Systems (RSS) . arXiv:2402.19249

    Pith/arXiv arXiv 2024

  30. [30]

    Tao Chen, Megha Tippur, Siyang Wu, Vikash Kumar, Edward Adelson, and Pulkit Agrawal. 2023. Visual Dexterity: In-Hand Reorientation of Novel and Complex Object Shapes. Science Robotics (2023). arXiv:2211.11744

    Pith/arXiv arXiv 2023

  31. [31]

    Wentse Chen, Shiyu Huang, Yuan Chiang, Ting Chen, and Jun Zhu. 2024. DGPO: Discovering Multiple Strate- gies with Diversity-Guided Policy Optimization. In Proceedings of the AAAI Conference on Artificial Intelligence . arXiv:2207.05631

    Pith/arXiv arXiv 2024

  32. [32]

    Yongchao Chen, Jacob Arkin, Yang Zhang, Nicholas Roy, and Chuchu Fan. 2023. AutoTAMP: Autoregressive Task and Motion Planning with LLMs as Translators and Checkers. arXiv preprint arXiv:2306.06531 (2023)

    Pith/arXiv arXiv 2023

  33. [33]

    Yongchao Chen, Rujul Gandhi, Yang Zhang, and Chuchu Fan. 2023. NL2TL: Transforming Natural Languages to Temporal Logics using Large Language Models. arXiv preprint arXiv:2305.07766 (2023)

    Pith/arXiv arXiv 2023

  34. [34]

    Zoey Chen, Sho Kiami, Abhishek Gupta, and Vikash Kumar. 2023. GenAug: Retargeting behaviors to unseen situations via Generative Augmentation. arXiv preprint arXiv:2302.06671 (2023)

    Pith/arXiv arXiv 2023

  35. [35]

    Xuxin Cheng, Yandong Ji, Junming Chen, Ruihan Yang, Ge Yang, and Xiaolong Wang. 2024. Expressive Whole-Body Control for Humanoid Robots. In Robotics: Science and Systems (RSS) . arXiv:2402.16796

    Pith/arXiv arXiv 2024

  36. [36]

    Xuxin Cheng, Kexin Shi, Ananye Agarwal, and Deepak Pathak. 2024. Extreme Parkour with Legged Robots. In IEEE International Conference on Robotics and Automation (ICRA) . arXiv:2309.14341

    Pith/arXiv arXiv 2024

  37. [37]

    Cheng Chi et al. 2023. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. In Robotics: Science and Systems (RSS). arXiv:2303.04137

    Pith/arXiv arXiv 2023

  38. [38]

    Cheng Chi et al. 2024. Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots. In Robotics: Science and Systems (RSS) . arXiv:2402.10329. Weights or Skills? 1:23

    Pith/arXiv arXiv 2024

  39. [39]

    Kurtland Chua, Roberto Calandra, Rowan McAllister, and Sergey Levine. 2018. Deep Reinforcement Learning in a Handful of Trials using Probabilistic Dynamics Models. In Advances in Neural Information Processing Systems (NeurIPS). arXiv:1805.12114

    Pith/arXiv arXiv 2018

  40. [40]

    Sudeep Dasari, Frederik Ebert, Stephen Tian, Suraj Nair, Bernadette Bucher, Karl Schmeckpeper, Siddharth Singh, Sergey Levine, and Chelsea Finn. 2019. RoboNet: Large-Scale Multi-Robot Learning. In Conference on Robot Learning (CoRL). arXiv:1910.11215

    Pith/arXiv arXiv 2019

  41. [41]

    Jingtao Ding et al. 2025. Understanding World or Predicting Future? A Comprehensive Survey of World Models. Comput. Surveys (2025). arXiv:2411.14499

    arXiv 2025

  42. [42]

    Yan Ding, Xiaohan Zhang, Saeid Amiri, Nieqing Cao, Hao Yang, et al . 2022. Robot Task Planning and Situation Handling in Open Worlds. arXiv preprint arXiv:2210.01287 (2022)

    Pith/arXiv arXiv 2022

  43. [43]

    Yan Ding, Xiaohan Zhang, Chris Paxton, and Shiqi Zhang. 2023. Task and Motion Planning with Large Language Models for Object Rearrangement. arXiv preprint arXiv:2303.06247 (2023)

    Pith/arXiv arXiv 2023

  44. [44]

    Ria Doshi, Homer Walke, Oier Mees, Sudeep Dasari, and Sergey Levine. 2024. Scaling Cross-Embodied Learn- ing: One Policy for Manipulation, Navigation, Locomotion and Aviation. In Conference on Robot Learning (CoRL) . arXiv:2408.11812

    Pith/arXiv arXiv 2024

  45. [45]

    Danny Driess et al. 2023. PaLM-E: An Embodied Multimodal Language Model. In International Conference on Machine Learning (ICML). arXiv:2303.03378

    Pith/arXiv arXiv 2023

  46. [46]

    Yilun Du et al. 2023. Learning Universal Policies via Text-Guided Video Generation. InAdvances in Neural Information Processing Systems (NeurIPS). arXiv:2302.00111

    Pith/arXiv arXiv 2023

  47. [47]

    Yuqing Du, Olivia Watkins, Zihan Wang, Cédric Colas, Trevor Darrell, Pieter Abbeel, Abhishek Gupta, and Jacob Andreas. 2023. Guiding Pretraining in Reinforcement Learning with Large Language Models. arXiv preprint arXiv:2302.06692 (2023)

    Pith/arXiv arXiv 2023

  48. [48]

    Jiafei Duan, Wilbert Pumacay, Nishanth Kumar, Yi Ru Wang, Shulin Tian, Wentao Yuan, Ranjay Krishna, Dieter Fox, Ajay Mandlekar, and Yijie Guo. 2024. AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation. arXiv preprint (2024). arXiv:2410.00371

    Pith/arXiv arXiv 2024

  49. [49]

    Ben Eisner, Harry Zhang, and David Held. 2022. FlowBot3D: Learning 3D Articulation Flow to Manipulate Articulated Objects. In Robotics: Science and Systems (RSS) . arXiv:2205.04382

    Pith/arXiv arXiv 2022

  50. [50]

    Benjamin Eysenbach, Abhishek Gupta, Julian Ibarz, and Sergey Levine. 2019. Diversity is All You Need: Learning Skills without a Reward Function. In International Conference on Learning Representations (ICLR) . arXiv:1802.06070

    Pith/arXiv arXiv 2019

  51. [51]

    Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. 2023. RH20T: A Comprehensive Robotic Dataset for Learning Diverse Skills in One-Shot. arXiv preprint arXiv:2307.00595 (2023)

    Pith/arXiv arXiv 2023

  52. [52]

    Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. 2023. AnyGrasp: Robust and Efficient Grasp Perception in Spatial and Temporal Domains.IEEE Transactions on Robotics (T-RO) (2023). arXiv:2212.08333

    Pith/arXiv arXiv 2023

  53. [53]

    Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. 2020. GraspNet-1Billion: A Large-Scale Bench- mark for General Object Grasping. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . doi:10.1109/CVPR42600.2020.01146

    arXiv 2020

  54. [54]

    Pete Florence et al. 2021. Implicit Behavioral Cloning. In Conference on Robot Learning (CoRL) . arXiv:2109.00137

    Pith/arXiv arXiv 2021

  55. [55]

    Carlos Florensa, Yan Duan, and Pieter Abbeel. 2017. Stochastic Neural Networks for Hierarchical Reinforcement Learning. In International Conference on Learning Representations (ICLR) . arXiv:1704.03012

    Pith/arXiv arXiv 2017

  56. [56]

    Kevin Frans, Jonathan Ho, Xi Chen, Pieter Abbeel, and John Schulman. 2018. Meta Learning Shared Hierarchies. In International Conference on Learning Representations (ICLR) . arXiv:1710.09767

    Pith/arXiv arXiv 2018

  57. [57]

    Justin Fu, Aviral Kumar, Ofir Nachum, George Tucker, and Sergey Levine. 2020. D4RL: Datasets for Deep Data-Driven Reinforcement Learning. arXiv preprint arXiv:2004.07219 (2020)

    Pith/arXiv arXiv 2020

  58. [58]

    Letian Fu, Justin Yu, Karim El-Refai, Ethan Kou, Haoru Xue, Huang Huang, Wenli Xiao, Guanzhi Wang, Dantong Niu, Li Fei-Fei, Guanya Shi, Jiajun Wu, Shankar Sastry, Yuke Zhu, Ken Goldberg, and Linxi Fan. 2026. CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation. arXiv preprint (2026). arXiv:2603.22435

    Pith/arXiv arXiv 2026

  59. [59]

    Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. 2024. HumanPlus: Humanoid Shadowing and Imitation from Humans. In Conference on Robot Learning (CoRL) . arXiv:2406.10454

    Pith/arXiv arXiv 2024

  60. [60]

    Zhao, and Chelsea Finn

    Zipeng Fu, Tony Z. Zhao, and Chelsea Finn. 2024. Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation. arXiv preprint arXiv:2401.02117 (2024)

    Pith/arXiv arXiv 2024

  61. [61]

    Scott Fujimoto and Shixiang Shane Gu. 2021. A Minimalist Approach to Offline Reinforcement Learning. In Advances in Neural Information Processing Systems (NeurIPS) . arXiv:2106.06860

    Pith/arXiv arXiv 2021

  62. [62]

    Scott Fujimoto, David Meger, and Doina Precup. 2019. Off-Policy Deep Reinforcement Learning without Exploration. In International Conference on Machine Learning (ICML) . arXiv:1812.02900. 1:24 Jena et al

    Pith/arXiv arXiv 2019

  63. [63]

    Samir Yitzhak Gadre, Mitchell Wortsman, Gabriel Ilharco, Ludwig Schmidt, and Shuran Song. 2023. CoWs on Pasture: Baselines and Benchmarks for Language-Driven Zero-Shot Object Navigation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . arXiv:2203.10421

    Pith/arXiv arXiv 2023

  64. [64]

    Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, et al. 2020. ThreeDWorld: A Platform for Interactive Multi-Modal Physical Simulation. arXiv preprint arXiv:2007.04954 (2020)

    Pith/arXiv arXiv 2020

  65. [65]

    Gemini Robotics Team, Google DeepMind. 2025. Gemini Robotics: Bringing AI into the Physical World.arXiv preprint (2025). arXiv:2503.20020

    Pith/arXiv arXiv 2025

  66. [66]

    Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. 2024. Octo: An Open-Source Generalist Robot Policy. In Robotics: Science and Systems (RSS). Octo Mode...

    Pith/arXiv arXiv 2024

  67. [67]

    Ankit Goyal et al. 2023. RVT: Robotic View Transformer for 3D Object Manipulation. In Conference on Robot Learning (CoRL). arXiv:2306.14896

    Pith/arXiv arXiv 2023

  68. [68]

    Ankit Goyal et al. 2024. RVT-2: Learning Precise Manipulation from Few Demonstrations. In Robotics: Science and Systems (RSS). arXiv:2406.08545

    Pith/arXiv arXiv 2024

  69. [69]

    Kristen Grauman et al. 2022. Ego4D: Around the World in 3,000 Hours of Egocentric Video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . arXiv:2110.07058

    Pith/arXiv arXiv 2022

  70. [70]

    Karol Gregor, Danilo Jimenez Rezende, and Daan Wierstra. 2016. Variational Intrinsic Control. arXiv preprint arXiv:1611.07507 (2016)

    Pith/arXiv arXiv 2016

  71. [71]

    Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. 2023. ManiSkill2: A Unified Benchmark for Generalizable Manipulation Skills. InInternational Conference on Learning Representations (ICLR). arXiv:2302.04659

    Pith/arXiv arXiv 2023

  72. [72]

    Ping Guo, Chao Li, Yinglan Feng, and Chaoning Zhang. 2026. Code Evolution for Control: Synthesizing Policies via LLM-Driven Evolutionary Search. arXiv preprint (2026). arXiv:2601.06845

    arXiv 2026

  73. [73]

    Yanjiang Guo, Yen-Jen Wang, Lihan Zha, and Jianyu Chen. 2024. DoReMi: Grounding Language Model by Detecting and Recovering from Plan-Execution Misalignment. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). arXiv:2307.00329

    Pith/arXiv arXiv 2024

  74. [74]

    Abhishek Gupta, Vikash Kumar, Corey Lynch, Sergey Levine, and Karol Hausman. 2019. Relay Policy Learning: Solving Long-Horizon Tasks via Imitation and Reinforcement Learning. In Conference on Robot Learning (CoRL) . arXiv:1910.11956

    Pith/arXiv arXiv 2019

  75. [75]

    Irmak Guzey, Yinlong Dai, Ben Evans, Soumith Chintala, and Lerrel Pinto. 2023. See to Touch: Learning Tactile Dexterity through Visual Incentives. arXiv preprint arXiv:2309.12300 (2023)

    Pith/arXiv arXiv 2023

  76. [76]

    Irmak Guzey, Ben Evans, Soumith Chintala, and Lerrel Pinto. 2023. Dexterity from Touch: Self-Supervised Pre-Training of Tactile Representations with Robotic Play. In Conference on Robot Learning (CoRL) . arXiv:2303.12076

    Pith/arXiv arXiv 2023

  77. [77]

    David Ha and Jürgen Schmidhuber. 2018. World Models. arXiv preprint arXiv:1803.10122 (2018)

    Pith/arXiv arXiv 2018

  78. [78]

    Huy Ha, Pete Florence, and Shuran Song. 2023. Scaling Up and Distilling Down: Language-Guided Robot Skill Acquisition. arXiv preprint arXiv:2307.14535 (2023)

    Pith/arXiv arXiv 2023

  79. [79]

    Danijar Hafner, Kuang-Huei Lee, Ian Fischer, and Pieter Abbeel. 2022. Deep Hierarchical Planning from Pixels. In Advances in Neural Information Processing Systems (NeurIPS) . arXiv:2206.04114

    Pith/arXiv arXiv 2022

  80. [80]

    Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. 2020. Dream to Control: Learning Behaviors by Latent Imagination. In International Conference on Learning Representations (ICLR) . arXiv:1912.01603

    Pith/arXiv arXiv 2020

Showing first 80 references.