Robotics · arXiv:2609.38178 v1 · submitted 29 Sep 2026 · 50 references
Skill-Space Shooting for Autonomous Robot Policy Improvement
The paper says
“Robots deployed in the physical world must be able to improve beyond their initial training as they encounter new situations and failures.”
more of the abstract
For this improvement to scale across tasks, it must make effective use of experience without requiring human demonstration of each correction. Recent agentic systems offer a way to reduce this reliance on human effort by using foundation models to autonomously compose learned behaviors to complete tasks. Yet completing tasks this way does not itself teach a task policy to overcome its own failures; that requires turning these behaviors into learnable corrections for the policy. Our insight is that many such corrections are familiar short behaviors, or skills: they recur across tasks and describe actions that foundation models can reason about from a scene. We introduce skill-space shooting, which uses foundation model guidance to explore corrections through these reusable skills and turn successful trials into policy improvement. Real-world experiments show repeated improvement in policies acting autonomously, while skills can also be shared to reduce the teaching needed to improve on new tasks. By making reusable skills a source of corrective supervision, skill-space shooting enables scalable and generalizable policy improvement within and across tasks. Additional results and videos at https://skill-space-shooting.github.io.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 30 Sep 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A clean mechanism — verified skill repairs as DAgger-style supervision, evaluated on the unaided policy — with honest experiments whose single-run statistics keep the headline numbers suggestive rather than measured. This line is wrong
3 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 50 references say.
The paper says, at Sec. 4, evaluation paragraph; b170; Table 1; Fig. 3
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The reported bootstrap 95% intervals resample only the 20–30 evaluation rollouts of a single run per method per task, so they quantify evaluation noise but cannot in principle capture run-to-run variance from finetuning seeds or rollout-collection randomness. The headline comparisons — OURS 0.600 vs DSRL 0.100 on Drawer (7/20 vs 2/20, b170), and the OURS-over-HUMAN gaps on Coffee (71.25 vs 46.25) and Sweeping (83.75 vs 63.13) — are single draws.
more of objection 1
The within-run improvement trends in Fig. 3 support the existence of improvement, but the displayed intervals imply a precision the design cannot support. Please either run multiple seeds per method per task, or restate all method-to-method comparisons as single-run observations and label the intervals explicitly as covering evaluation noise only.
The paper says, at Sec. 4.2; b187
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The claimed 97.3–98.8% repair probability is computed as 1−(1−p)^3 for the per-attempt success range p∈[0.70,0.77], which treats up to three retries as independent, identically distributed attempts. Retries of the same skill at the same failure state are likely correlated, since a shared root cause (e.g., a systematic grasp offset) can defeat every attempt; under such correlation the effective per-rollout rate stays near the 70–77% per-attempt figure.
more of objection 2
Report the empirically observed per-rollout repair rate under the retry budget instead of, or alongside, the independence-computed number, and do not present the latter as the achieved reliability.
The paper says, at Sec. 3.1; Table 2
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The authors acknowledge that each task stage has a designated primitive, which means the judge's 90–100% 'stage/skill accuracy' largely reflects stage identification rather than semantic skill selection. What remains untested is whether foundation-model semantic selection contributes anything over a non-semantic stage-progression heuristic (e.g., pick the next stage's skill by task order).
more of objection 3
Because the loop is otherwise identical in either case, the reported accuracy does not isolate the semantic component that the paper's framing emphasizes. A control that selects skills without scene reasoning would clarify whether the gains come from the shooting-and-retraining mechanism itself or from the semantic selection, and would strengthen the interpretation of Table 2.
5 minor objections
The paper says, at Sec. 6; b220
Pith says: minor objection 1 of 5
The sentence 'We also prove that sharing skill data reduces the new teaching needed...' uses 'prove' for an empirical demonstration. Suggest 'demonstrate' to avoid overclaiming.
The paper says, at General
Pith says: minor objection 2 of 5
No code, data, or policy/skill checkpoints are released. For a four-task real-robot study, releasing the trained policies, value models, and skill libraries would substantially aid reproducibility.
show 3 more
The paper says, at Table 1 footnote
Pith says: minor objection 3 of 5
The footnote 'a bound covering its 95% bootstrap CI' should state explicitly that the bound reflects evaluation noise for the single run and not run-to-run variance, consistent with the major comment above.
The paper says, at Sec. 4.1
Pith says: minor objection 4 of 5
The claim that 'our initial policies begin below COP, then exceed it on every task' would be clearer if the authors specified the iteration at which OURS overtakes COP per task, since COP is a single evaluation.
The paper says, at Sec. 4.3; b195; b196
Pith says: minor objection 5 of 5
For the Pick Sponge results (grasp success 5%→70% with 10 demos; the 30-demo matching claim), specify the evaluation protocol — number of trials, object placements, and whether grasp success is measured per attempt or per rollout — so the numbers are reproducible.
What would settle it
Re-run the full Drawer improvement loop (four rounds, 21-81 rollouts per round, as in Sec. 4) with at least five independent seeds of the pi0.5 finetuning, keeping the value model, judge, and retry budget fixed, and report the distribution of unaided success after update 1 and at the plateau. If 7/20 falls inside the seed-to-seed spread (e.g., across-seed successes span 3-9/20 and overlap the DSRL 2/20 outcome), the improvement claim weakens materially; if every seed rises above the initial 0/20, the claim holds. Repeat for one Coffee or Sweeping round to confirm the progress advances are not a single-run artifact.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 30 Sep 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
What you should know before reading it: the mechanism is real and unusually clean. A value model flags likely failures, a VLM selects a skill to shoot from the learner's current state, and accepted, judge-confirmed repair segments are aggregated into the task policy as DAgger-style supervision. Evaluation strips away the value model, judge, and skill library, so the reported gains are measured on the unaided policy — the right design choice, and it mostly defuses the circularity worry.
Credit where due. The HUMAN baseline is the important control: fresh full-task demonstrations matched in total duration to the retained repair data, updated with the same procedure, and OURS beat it on all four tasks. That is genuine evidence the improvement isn't just "more imitation data" — where the supervision is collected matters. The authors also fix one trigger threshold and one action horizon across tasks, openly state the single-run evaluation design, and list their limitations (scene resets, per-task detector training) without glossing. The sharing experiments include matched-new-demonstration internal controls. The supervision-validity question the stress-test raises is the right one to ask, and the HUMAN comparison is the best available answer short of a dedicated ablation.
Soft spots, in proportion. The load-bearing one is statistical: one run per method per task, with bootstrap intervals that resample only that run's 20-30 evaluation rollouts. Those CIs quantify evaluation noise within a single run; they cannot capture run-to-run variance from finetuning seeds and collection randomness. So the Drawer 0/20-to-7/20 jump, the OURS-over-HUMAN gaps on Coffee and Sweeping, and the sharing gains are single draws. The pattern is consistent across four tasks and the mechanism is internally coherent, which makes the claim credible, but any individual number is suggestive, not measured. Second: Section 3.1 admits each stage has a designated primitive, so the judge's 90-100% "skill selection" accuracy is essentially stage classification — honest, but it caps how much credit the VLM-selection story can take. Third: the retry arithmetic assumes independent attempts; retries from the same failure state are likely correlated, so the claimed 97-99% effective repair rate is probably optimistic. That inflates a supporting figure, not the core claim. Minor: the conclusion says "we prove" sharing reduces teaching — that's an empirical demonstration on a few tasks; loose wording. And there is no code or data release, which matters more here than usual given the single-run evidence.
Who it's for: anyone building autonomous improvement or self-correction loops for deployed robot policies. It deserves a serious referee. The right close would be to push for multi-seed runs, or at minimum error bars that honestly cover run-level uncertainty. I'd expect a conditional accept trajectory, not a desk reject.
The strongest objection
The central claim (b38, b41, b170) is that skill-space shooting improves an unaided task policy across iterations, including from zero initial successes. Everything that would have to be true for this to hold is plausibly present: the repair segments are collected at learner-visited states, only accepted repairs from judge-confirmed complete rollouts are retained (Sec. 3.2), and evaluation strips away the value model, judge, and library. The HUMAN baseline, which matches the total duration of retained repair data with fresh full-task demonstrations and is beaten by OURS on all four tasks, is real supporting evidence that the improvement is not merely 'more data' - per unit of data, the repairs are more effective. The author-stated limitations (b210, b213) are honestly scoped. The load-bearing weakness is statistical: every method has one run per task, and the reported bootstrap 95% intervals (Sec. 4) resample the 20-30 evaluation rollouts of that single run. A bootstrap over one run can quantify evaluation noise but cannot in principle quantify run-to-run (finetuning-seed, rollout-collection) variance, so the units of uncertainty the paper displays systematically understate the uncertainty around the central result. The 7/20 versus DSRL 2/20 comparison, the OURS-over-HUMAN gaps on Coffee and Sweeping, and the sharing gains in Fig. 4c are all single-draw observations. Relatedly, the retry arithmetic in Sec. 4.2 (b187) assumes retries are independent and identically distributed; retries launched from the same failure state are likely correlated, so the claimed 97.3-98.8% effective repair rate is probably optimistic, though this inflates a supporting number rather than the core claim. The reader's supervision-validity concern is real but is partially mitigated by the HUMAN comparison; the single-run evidence is the least-secure load-bearing point.
The reader's weakest assumption - that skills supply valid, transferable corrective actions at the diverse failure states the policy reaches - is a genuine mechanism-level risk, and the Section 3.1 admission that each stage has a designated primitive does reduce the 90-100% 'skill selection' accuracy to stage classification. However, my reading is that this is not the single most load-bearing point: the improvement loop only needs the correct skill to fire at the correct time, and whether the choice comes from semantic reasoning or deterministic stage mapping does not change whether the policy improves. The HUMAN baseline (matched-duration full demonstrations, beaten by OURS on every task) already provides indirect but real evidence that the repair segments are effective supervision per data unit, partially answering the validity worry. What no baseline and no reported number addresses is run-to-run variance: one run per method per task, with bootstrap CIs computed by resampling the evaluation rollouts of that single run. Those intervals measure evaluation noise, not the training-seed and collection randomness that dominate in real-robot finetuning, so the uncertainty around the central claim (0/20 to 7/20; OURS exceeding DSRL, COP, and HUMAN) is systematically understated. This is concrete, testable, and decisive for whether the central claim is a robust empirical contribution or a favorable single draw. I agree with the reader's CONDITIONAL verdict: the framework is plausible and internally consistent, the author-stated limitations are honest, and the missing artifacts and statistics are the blocking items. My concern reinforces the conditions rather than overturning the verdict, so I recommend no change. Agreement with the reader is partial: we converge on CONDITIONAL, but I weight the single-run statistics over the supervision-validity assumption as the most load-bearing gap.
The referee's report
If the reported results hold, the paper makes a genuine contribution: it turns reusable short-horizon skills into autonomous corrective supervision, and its evaluation is set up in the right way to support that claim. The final evaluation strips away the value model, judge, and skill library, so the headline gains are measured against a non-circular benchmark; the HUMAN baseline, which matches the total duration of retained repair data with fresh full-task demos and is beaten by OURS on all four tasks, is real evidence that the improvement is not merely 'more data'; and the cost is honestly reported (only scene resets beyond initial teaching). The acknowledged limitations (per-task value-model training, human scene resets, the stage-designated-primitive simplification) are stated clearly. At the same time, the experimental evidence is thinner than the text suggests: every method has one run per task, and the displayed bootstrap intervals quantify only evaluation noise within that single run, not run-to-run variance.
Pith's reading of the paper (a machine's summary, not the authors')
The paper tries to establish that short, reusable robot behaviors — skills such as picking, wiping, and placing — can be invoked autonomously to correct a task policy's failures, and that the successful corrections can then be turned into training data that improves the policy without new human demonstrations. The method, skill-space shooting, runs an imperfect policy, uses a frozen value model to flag likely failures, and uses a foundation-model judge to choose which skill to try and to verify whether the repair worked. Only accepted, skill-executed repair segments from completed rollouts are kept and aggregated into the policy through DAgger-style updates. Across four real-world manipulation tasks the policy's unaided performance improves over successive rounds, including a Drawer policy with zero observed initial successes improving from 0/20 to 7/20 after one update, and sharing skill data across tasks reduces the new teaching needed. The claim matters because it proposes a route to closed-loop, self-generated corrective supervision that does not bottleneck on human correction of every failure.
What carries the argument. The mechanism is a three-part autonomous improvement loop. A value model $V_\phi$ scores the current observation together with the policy's proposed action chunk and triggers a shooting trial when the value drops below a fixed threshold; a foundation-model judge selects, from a library of learned primitives, which skill to execute and later checks whether the stage was repaired; and accepted repair segments from judge-confirmed completed rollouts are aggregated into the policy's training set in DAgger style, $D_{k+1} = D_0 \cup \bigcup_{j=0}^k R_j$. The central object is the skill library itself, whose identity (visual and textual descriptions) makes each candidate correction intelligible to the judge while its policy supplies the coordinated actions.
Core claim. The central discovery is that a library of separately taught short-horizon skills can act as a renewable source of corrective supervision for a long-horizon task policy. Because each skill only needs to solve one local stage, it can supply useful actions at the failure states the policy actually reaches — states that a fresh demonstration from a clean starting position may never cover. The loop detects when help is needed with a value model, lets a foundation-model judge select and verify a skill repair, and trains the policy on only the accepted repair segments, which isolates the correction from the policy's own mistakes. The paper reports that this improves unaided performance on all four tasks over successive iterations, can improve a policy with no observed initial successes, and that sharing primitive data across tasks cuts the demonstrations needed to prepare corrections for a new task.
The premise it rests on. A skill trained for one stage of a task supplies correct, transferable actions at the varied failure states the task policy reaches, and keeping only accepted repair segments cleanly separates those corrections from the policy's own mistakes.
If the paper is right (Pith's inferences, not the authors')
A task policy can improve even when it starts with zero observed successes: the Drawer policy went from 0/20 to 7/20 after one update, because targeted local repairs let it construct successful attempts it could then learn from.
The gains persist when the value model, judge, and library are removed at evaluation, so the corrective behavior is genuinely internalized by the task policy rather than borrowed at test time.
Skill-space shooting reaches a higher plateau than latent-noise RL (DSRL), direct primitive composition (COP), and additional end-to-end human demonstrations (HUMAN) on all four tasks, with the largest margins on tasks requiring sustained contact or long sequences.
Sharing primitive training data across tasks reduces the demonstrations needed for a new task: with only 10 new sponge demonstrations, sharing raises grasp success from 5% to 70%, and a shared library lets a new task reach full observed success one iteration earlier.
Because the paper notes each task stage has a designated primitive, the judge's 90–100% skill-selection accuracy is largely stage identification; a testable extension is to let multiple skills compete for a stage and check whether the semantic choice adds anything beyond detecting which stage failed.
The value model is trained once on initial labeled episodes and frozen, so the loop cannot adapt its failure detection to the policy's shifting error modes over iterations; retraining the detector each round is a natural extension the paper leaves implicit.
The two-task sharing result hints at a compounding return in larger libraries, where each newly taught primitive lowers the teaching cost for many later tasks; quantifying that compounding across more than two tasks would test whether the library genuinely scales.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Robots deployed in the physical world must be able to improve beyond their initial training as they encounter new situations and failures. For this improvement to scale across tasks, it must make effective use of experience without requiring human demonstration of each correction. Recent agentic systems offer a way to reduce this reliance on human effort by using foundation models to autonomously compose learned behaviors to complete tasks. Yet completing tasks this way does not itself teach a task policy to overcome its own failures; that requires turning these behaviors into learnable corrections for the policy. Our insight is that many such corrections are familiar short behaviors, or skills: they recur across tasks and describe actions that foundation models can reason about from a scene. We introduce skill-space shooting, which uses foundation model guidance to explore corrections through these reusable skills and turn successful trials into policy improvement. Real-world experiments show repeated improvement in policies acting autonomously, while skills can also be shared to reduce the teaching needed to improve on new tasks. By making reusable skills a source of corrective supervision, skill-space shooting enables scalable and generalizable policy improvement within and across tasks. Additional results and videos at https://skill-space-shooting.github.io.
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 · 2026-09-30
The authors' main claim, as the classifier picked it out 0.99
“By making reusable skills a source of corrective supervision, skill-space shooting enables scalable and generalizable policy improvement within and across tasks.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.14 · human participants 0.10 · controlled comparison 0.08 · randomized assignment 0.01 · states quantitative result 0.04 · compares with prior results 0.15 · mentions code or data availability 0.50
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Robots deployed in the physical world must be able to improve beyond their initial training as they encounter new situations and failures. | background 1.00 | 0.05 | 0.03 | asserted 0.94 |
| For this improvement to scale across tasks, it must make effective use of experience without requiring human demonstration of each correction. | background 0.93 | 0.05 | 0.03 | asserted 0.94 |
| Recent agentic systems offer a way to reduce this reliance on human effort by using foundation models to autonomously compose learned behaviors to complete tasks. | background 1.00 | 0.04 | 0.03 | asserted 0.74 |
| Yet completing tasks this way does not itself teach a task policy to overcome its own failures; that requires turning these behaviors into learnable corrections for the policy. | gap 0.96 | 0.07 | 0.04 | asserted 1.00 |
| Our insight is that many such corrections are familiar short behaviors, or skills: they recur across tasks and describe actions that foundation models can reason about from a scene. | contribution 0.53 | 0.36 | 0.08 | asserted 0.97 |
| We introduce skill-space shooting, which uses foundation model guidance to explore corrections through these reusable skills and turn successful trials into policy improvement. | contribution 0.60 | 0.85 | 0.16 | asserted 1.00 |
| Real-world experiments show repeated improvement in policies acting autonomously, while skills can also be shared to reduce the teaching needed to improve on new tasks. | result 1.00 | 0.95 | 0.25 | asserted 1.00 |
| By making reusable skills a source of corrective supervision, skill-space shooting enables scalable and generalizable policy improvement within and across tasks. | interpretation 0.91 | 0.79 | 0.84 | asserted 1.00 |
| Additional results and videos at https://skill-space-shooting.github.io. | future work 0.66 | 0.17 | 0.02 | unclear 0.88 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 1.00
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, 2026-09-30: 1 request, 6,810 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (4)
Figures are the authors', reproduced from the paper.
References (50)
- [1] A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning , author =. Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics (AISTATS) , series =. 2011 , url =. 1011.0686 , archivePrefix = on PithPith's reviewarXiv
- [2] Hoque, Ryan and Balakrishna, Ashwin and Novoseller, Ellen and Wilcox, Albert and Brown, Daniel S. and Goldberg, Ken , booktitle =. 2021 , url =. 2109.08273 , archivePrefix = on PithPith's reviewarXiv
- [3] Human-in-the-Loop Imitation Learning using Remote Teleoperation , author =. arXiv preprint arXiv:2012.06733 , year =. 2012.06733 , archivePrefix = on PithPith's reviewarXiv
- [4] Robot Learning on the Job: Human-in-the-Loop Autonomy and Learning During Deployment , author =. Robotics: Science and Systems (RSS) , year =. doi:10.15607/RSS.2023.XIX.005 , eprint = on Pithdoi:10.15607/rss.2023.xix.005
- [5] Shi, Lucy Xiaoyang and Hu, Zheyuan and Zhao, Tony Z. and Sharma, Archit and Pertsch, Karl and Luo, Jianlan and Levine, Sergey and Finn, Chelsea , booktitle =. 2024 , doi =. 2403.12910 , archivePrefix = on PithPith's reviewarXiv
- [6] Robot-Gated Interactive Imitation Learning with Adaptive Intervention Mechanism , author =. Proceedings of the 42nd International Conference on Machine Learning (ICML) , series =. 2025 , url =. 2506.09176 , archivePrefix = on PithPith's reviewarXiv
- [7] Distilling and Retrieving Generalizable Knowledge for Robot Manipulation via Language Corrections , author =. IEEE International Conference on Robotics and Automation (ICRA) , pages =. 2024 , doi =. 2311.10678 , archivePrefix = on PithPith's reviewarXiv
- [8] To Err is Robotic: Rapid Value-Based Trial-and-Error during Deployment , author =. arXiv preprint arXiv:2406.15917 , year =. 2406.15917 , archivePrefix = on PithPith's reviewarXiv
- [9] Lin, Zijun and Duan, Jiafei and Fang, Haoquan and Fox, Dieter and Krishna, Ranjay and Tan, Cheston and Wen, Bihan , journal =. 2025 , eprint = on Pith
- [10] Zhang, Fan and Park, Seongbin and Mirzasoleiman, Baharan and Talebi, Shahriar and Sehatbakhsh, Nader , journal =. 2026 , eprint = on Pith
- [11] Recover, Discover, Plan: Learning Skills and Concepts from Robot Failures , author =. arXiv preprint arXiv:2606.18328 , year =. 2606.18328 , archivePrefix = on PithPith's reviewarXiv
- [12] Deng, Haoyuan and Gao, Yitong and Lin, Yudong and Liu, Haichao and Wu, Zhenyu and Wang, Ziwei , journal =. 2026 , eprint = on Pith
- [13] Yi, Pengfei and Ma, Yingjie and Xu, Wenjiang and Hao, Yanan and Gan, Shuai and Li, Wanting and Zhong, Shanlin , journal =. 2026 , eprint = on Pith
- [14] Sa, Inkyu and Stulov, Konstantin and Bhageria, Rajat , journal =. 2026 , eprint = on Pith
- [15] Hu, Haodi and Huang, Chung-Ta and Liu, Jing and Wang, Ye and Suzuki, Kei and Brand, Matthew and Koike-Akino, Toshiaki , journal =. 2026 , eprint = on Pith
- [16] Zhao, Ganlong and Tang, Zijia and Chen, Xingping and Kuang, Zhanghui and Tian, Ye and Li, Guanbin , booktitle =. 2026 , url =. 2608.26645 , archivePrefix = on PithPith's reviewarXiv
- [17] Do As I Can, Not As I Say: Grounding Language in Robotic Affordances , author =. Proceedings of the 6th Conference on Robot Learning (CoRL) , series =. 2022 , url =. 2204.01691 , archivePrefix = on PithPith's reviewarXiv
- [18] Huang, Haoxu and Lin, Fanqi and Hu, Yingdong and Wang, Shengjie and Gao, Yang , booktitle =. 2024 , doi =. 2403.08248 , archivePrefix = on PithPith's reviewarXiv
- [19] Shi, Lucy Xiaoyang and Ichter, Brian and Equi, Michael Robert and Ke, Liyiming and Pertsch, Karl and Vuong, Quan and Tanner, James and Walling, Anna and Wang, Haohuan and Fusai, Niccolo and Li-Bell, Adrian and Driess, Danny and Groom, Lachy and Levine, Sergey and Finn, Chelsea... on PithPith's reviewarXiv
- [20] Belkhale, Suneel and Ding, Tianli and Xiao, Ted and Sermanet, Pierre and Vuong, Quan and Tompson, Jonathan and Chebotar, Yevgen and Dwibedi, Debidatta and Sadigh, Dorsa , booktitle =. 2024 , doi =. 2403.01823 , archivePrefix = on PithPith's reviewarXiv
show 30 more
- [21] Liang, Jacky and Huang, Wenlong and Xia, Fei and Xu, Peng and Hausman, Karol and Ichter, Brian and Florence, Pete and Zeng, Andy , booktitle =. 2023 , doi =. 2209.07753 , archivePrefix = on PithPith's reviewarXiv
- [22] Programmatically Grounded, Compositionally Generalizable Robotic Manipulation , author =. International Conference on Learning Representations (ICLR) , year =. 2304.13826 , archivePrefix = on PithPith's reviewarXiv
- [23] Wang, Chen and Fan, Linxi and Sun, Jiankai and Zhang, Ruohan and Fei-Fei, Li and Xu, Danfei and Zhu, Yuke and Anandkumar, Anima , booktitle =. 2023 , url = on Pith
- [24] Ding, Zi-Han and Wang, Ziwei , journal =. 2026 , doi =. 2607.08354 , archivePrefix = on PithPith's reviewarXiv
- [25] Li, Ruiying and Zhou, Yunlang and Zhu, YuYao and Chen, Kylin and Wang, Jingyuan and Wang, Sukai and Hu, Kongtao and Yu, Minhui and Jiang, Bowen and Su, Zhan and Ma, Jiayao and He, Xin and Shen, Yongjian and Yang, Yang and Ren, Guanghui and Yao, Maoqing and Wang, Wenhao and Mu,... on Pith
- [26] Lu, Runyu and Wu, Yubo and Kou, Ethan and Fu, Letian and Xiao, Wenli and Mandlekar, Ajay and Xu, Yinzhen and Shi, Guanya and Goldberg, Ken and Chen, Ang and Chowdhury, Mosharaf and Zhu, Yuke and Fan, Linxi and Wang, Guanzhi , journal =. 2026 , eprint = on Pith
- [27] Bousmalis, Konstantinos and Vezzani, Giulia and Rao, Dushyant and Devin, Coline and Lee, Alex X. and Bauza, Maria and Davchev, Todor and Zhou, Yuxiang and Gupta, Agrim and Raju, Akhil and Laurens, Antoine and Fantacci, Claudio and Dalibard, Valentin and Zambelli, Martina and M... on PithPith's reviewarXiv
- [28] Autonomous Improvement of Instruction Following Skills via Foundation Models , author =. Proceedings of the 8th Conference on Robot Learning (CoRL) , series =. 2024 , url =. 2407.20635 , archivePrefix = on PithPith's reviewarXiv
- [29] arXiv preprint arXiv:2511.14759 , year =. 2511.14759 , archivePrefix = on PithPith's reviewarXiv
- [30] Giridhar, Varun and Khandelwal, Anant and Collins, Jeremy A. and Georgiev, Ignat and Garg, Animesh , journal =. Beyond Imitation: Self-Improving Robot Policies via Off-Policy. 2026 , eprint = on Pith
- [31] Hu, Zheyuan and Wu, Robyn and Enock, Naveen and Li, Jasmine and Kadakia, Riya and Erickson, Zackory and Kumar, Aviral , journal =. 2025 , eprint = on Pith
- [32] Xiao, Wenli and Lin, Haotian and Peng, Andy and Xue, Haoru and He, Tairan and Xie, Yuqi and Hu, Fengyuan and Wu, Jimmy and Luo, Zhengyi and Fan, Linxi and Shi, Guanya and Zhu, Yuke , booktitle =. Self-Improving Vision-Language-Action Models with Data Generation via Residual. 2... on PitharXiv
- [33] Self-Improving Robot Policy with Compositional World Model , author =. Robotics: Science and Systems (RSS) , year =. doi:10.15607/RSS.2026.XXII.012 , eprint = on Pithdoi:10.15607/rss.2026.xxii.012
- [34] So You Think You Can Scale Up Autonomous Robot Data Collection? , author =. Proceedings of the 8th Conference on Robot Learning (CoRL) , series =. 2024 , url =. 2411.01813 , archivePrefix = on PithPith's reviewarXiv
- [35] Data Quality in Imitation Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , volume =. 2023 , doi =. 2306.02437 , archivePrefix = on PithPith's reviewarXiv
- [36] Shumailov, Ilia and Shumaylov, Zakhar and Zhao, Yiren and Papernot, Nicolas and Anderson, Ross and Gal, Yarin , journal =. 2024 , doi = on Pith
- [37] Black, Kevin and Brown, Noah and Darpinian, James and Dhabalia, Karan and Driess, Danny and Esmail, Adnan and Equi, Michael Robert and Finn, Chelsea and Fusai, Niccolo and Galliker, Manuel Y. and Ghosh, Dibya and Groom, Lachy and Hausman, Karol and Ichter, Brian and Jakubczak,... on PithPith's reviewarXiv
- [38] Steering Your Diffusion Policy with Latent Space Reinforcement Learning , author =. Proceedings of the 9th Conference on Robot Learning (CoRL) , series =. 2025 , url =. 2506.15799 , archivePrefix = on PithPith's reviewarXiv
- [39] Chi, Cheng and Xu, Zhenjia and Pan, Chuer and Cousineau, Eric and Burchfiel, Benjamin and Feng, Siyuan and Tedrake, Russ and Song, Shuran , booktitle =. 2024 , doi =. 2402.10329 , archivePrefix = on PithPith's reviewarXiv
- [40] Oquab, Maxime and Darcet, Timoth. Transactions on Machine Learning Research , year =. 2304.07193 , archivePrefix = on PithPith's reviewarXiv
- [41] arXiv preprint arXiv:2507.06261 , year =. 2507.06261 , archivePrefix = on PithPith's reviewarXiv
- [42] Flow Matching for Generative Modeling , author =. International Conference on Learning Representations (ICLR) , year =. 2210.02747 , archivePrefix = on PithPith's reviewarXiv
- [43] Black, Kevin and Brown, Noah and Driess, Danny and Esmail, Adnan and Equi, Michael Robert and Finn, Chelsea and Fusai, Niccolo and Groom, Lachy and Hausman, Karol and Ichter, Brian and Jakubczak, Szymon and Jones, Tim and Ke, Liyiming and Levine, Sergey and Li-Bell, Adrian and... on PithPith's reviewarXiv
- [44] Sutton, Richard S. and Precup, Doina and Singh, Satinder , journal =. Between. 1999 , doi = on Pith
- [45] Learning Complex Neural Network Policies with Trajectory Optimization , author =. Proceedings of the 31st International Conference on Machine Learning , pages =. 2014 , editor = on Pith
- [46] Accelerating Reinforcement Learning with Learned Skill Priors , author =. Proceedings of the 2020 Conference on Robot Learning , pages =. 2021 , editor = on Pith
- [47] Demonstration-Guided Reinforcement Learning with Learned Skills , author =. Proceedings of the 5th Conference on Robot Learning , pages =. 2022 , editor = on Pith
- [48] Kahn, Gregory and Zhang, Tianhao and Levine, Sergey and Abbeel, Pieter , booktitle =. 2017 , eprint = on Pith
- [49] Gupta, Abhishek and Yu, Justin and Zhao, Tony Z. and Kumar, Vikash and Rovinsky, Aaron and Xu, Kelvin and Devlin, Thomas and Levine, Sergey , title =. IEEE International Conference on Robotics and Automation , year = on Pith
- [50] Hansen, Steven and Xu, Peng , title =. 2026 , month = jul, howpublished = on Pith
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment