Robotics · arXiv:2609.34261 v1 · submitted 28 Sep 2026 · 16 references
RoboICL: Embodied In-Context Learning with GPT-6 Astra
The paper says
“General-purpose vision-language models offer a promising way to zero-shot robot control: \gptastra excels at open-ended and language- or image-conditioned manipulation but remains substantially weaker on high-precision and long-horizon tasks.”
more of the abstract
We introduce RoboICL, an in-context robot-control framework that narrows these gaps without robot-specific parameter updates or a learned VLA. RoboICL separates demonstration context, which provides recorded examples when available, from interaction memory, which accumulates the model's own actions and observed outcomes. Both use a shared observation–action–receipt–observation grammar. To preserve experience across task stages, RoboICL combines sampled demonstration blocks with bounded anchored memory. Fixed anchors keep earlier rollout interactions available for in-context learning, while the latest interaction supports immediate error correction. Across 30 RoboDojo tasks, using zero shot for Open and one demonstration elsewhere, RoboICL improves on official zero-shot \gptastra by 20–27 progress-score points in every category. It leads the leaderboard baselines on Memory and Open, achieves comparable performance to the strongest Precision baseline, and remains competitive on Long-Horizon. Its 30-task Overall score is 50.64, versus 33.68 for the strongest baseline. On a separate ten-task subset, RoboICL scores 60.60, within 2.00 points of the π0.5 + \gptastra hybrid approach. On three real-robot tasks, mean progress rises from 14.45 at zero shot to 63.33 at one shot and 78.89 at three shots. On two development tasks, optional Jev-gated action reuse reduces \gptastra calls by 33–48%. Code is available at \hrefhttps://github.com/Mosi-AI/RoboICLhttps://github.com/Mosi-AI/RoboICL.
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 transparent, well-engineered frozen-VLM ICL framework with plausible headline numbers that aren't yet verified — single-seed runs and dev-set-tuned allocations make a multi-seed rerun the gating condition for accept. 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 16 references say.
The paper says, at §4.2, Table 2
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
All RoboICL evaluations use seed 0 with no variance reported, while the VLA/WAM baselines in Table 2 average three seeds over 50 episodes per task. RoboProbe's own per-task scores show large spread within categories (e.g., Precision: Build Tower 16.40 and Insert Tubes 6.00 versus 82–87 for Liber-0 and Simate-beta), so the 20–27-point category gains and the 16.96-point Overall lead could be inflated if seed 0 is favorable or if the single-seed official baseline is weak on those tasks.
more of objection 1
Because the abstract's headline claim is a quantified margin, a multi-seed rerun (or at least per-layout score distributions) is needed before the margins can be taken as representative rather than seed-dependent.
The paper says, at §4.1, Fig 4
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.84):
“Because these tasks and layouts also informed context allocation, validation on held-out tasks and layouts remains necessary to establish generality.”
Pith says: Objection 2 of 3 · major
The main configuration was selected on four development tasks — Build Tower, Cover Blocks, Put Bottles into Dustbin, Insert Tubes — all of which reappear in the Table 2 category means (Precision, Memory, Long-Horizon, Precision respectively). Both the balanced allocation J=B=12 and the anchored-versus-recent memory choice come from this set, and the authors acknowledge that held-out validation is necessary (b149).
more of objection 2
The Overall score and category leads are therefore measured on a benchmark that informed the hyperparameters; out-of-sample validation on tasks and layouts not used in selection is required before the claimed margins can be credited as general, rather than as tuned on the evaluation set.
The paper says, at §3.1, Fig 3
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The 20–27-point comparison to official zero-shot GPT-6 Astra is a system-level comparison, not an ablation of the in-context organization: RoboICL differs from the official controller in action parameterization (per-step Cartesian increments versus absolute grasp-point targets), horizon schedule, image packaging (triptych versus three separate attachments), and execution rate.
more of objection 3
The more controlled comparison against GPT-as-Policy Direct (Table 1) uses only five layouts per task and marks two one-shot entries as reused zero-shot runs (dagger), which limits its statistical power. The abstract and conclusion should state explicitly that the margins are against a different controller design, and that the contribution of the shared grammar and anchored memory is supported primarily by the internal ablations in Section 4.1 rather than by the leaderboard comparisons.
5 minor objections
The paper says, at Table 1
Pith says: minor objection 1 of 5
Clarify that the dagger-marked one-shot entries (Classify by Language, Imitate Sorting Sequence) reuse zero-shot runs, and that Classify by Language has no demonstration available so its one-shot value equals its zero-shot value by protocol.
The paper says, at Fig 4
Pith says: minor objection 2 of 5
The selection between (12,12) and (16,8) rests on a 0.25-point difference in the four-task means (80.25 versus 80.00); report per-layout scores or variance so the reader can see the spread that the choice is sensitive to.
show 3 more
The paper says, at §4.3, Table 3
Pith says: minor objection 3 of 5
The Folding Towel out-of-distribution result is a single condition on one larger towel; the 60-point drop is consistent with the authors' own caution (b227) but would be more informative with per-trial scores rather than the mean only.
The paper says, at Abstract / §4.2
Pith says: minor objection 4 of 5
Note in the abstract that the ten-task panel uses five requested layouts per task, versus 50 episodes for the official evaluation, so the 60.60 versus 62.60 comparison to the hybrid is across different episode counts.
The paper says, at §4.5, Table 5
Pith says: minor objection 5 of 5
The Jev reduction figures (33–48% fewer GPT-6 Astra calls) are computed on the two development layouts, as the authors state (b259); it would be useful to state in the main text that Jev's own gate calls and tokens are excluded from the reported GPT-6 Astra totals.
What would settle it
Rerun the main RoboICL zero-shot (B=25) and one-shot (J=B=12) configurations on all 30 RoboDojo tasks at seeds 1–3, holding all hyperparameters fixed and reporting per-task means, category means, and Overall with standard errors. Falsification thresholds: if any category's seed-1–3 mean gain over official zero-shot GPT-6 Astra drops below roughly 10 points, or the seed-1–3 Overall mean falls below about 45 (versus 50.64 at seed 0), the headline margins are seed-sensitive and the 'leads the leaderboard' claim needs qualification. Reading the same runs on the four development tasks also settles whether the dev-selected (J,B) allocation and anchored-memory choice transfer: if the four-task ordering between allocations is stable across seeds, the tuning concern is resolved; if the dev tasks' advantage shrinks or reverses at other seeds, the allocation selection was partly overfit to seed 0.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 30 Sep 2026
- Reader holds up, with conditions
- Skeptic 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
The short version: this is a solid, unusually transparent empirical study of in-context control with a frozen frontier VLM. The headline numbers are plausible but not yet verified — all RoboICL scores come from seed 0, and the context allocations were chosen on four development tasks that also appear in the headline benchmark. To the authors' credit, they say exactly this in the text (b149: validation on held-out tasks and layouts remains necessary), so the issue is magnitude, not awareness.
What is actually new: the shared observation-action-receipt-observation grammar binding demonstration context to interaction memory, bounded anchored memory with explicit TRAJECTORY_GAP records, and per-step Cartesian-increment action chunks with adaptive horizons. None of the cited concurrent work has that combination: GPT-as-Policy Direct uses absolute targets in a persistent chat, GPT-Policy uses absolute waypoints with controller-determined timing. The 30-task evaluation is broader than the concurrent panels, and the three real-robot tasks with an honest OOD failure analysis (the larger towel drops to 40) are genuinely useful. The engineering is well documented: deterministic selection rules in Appendix A.1, a byte-identical cache-frontier verification over 1,780 requests in Appendix C.1, and an explicit development-set disclosure for Jev (b259). The citations to the baselines and concurrent work look sound.
Soft spots, in order. First, seed 0: Table 2 states all RoboICL evaluations use seed 0 while baselines use three seeds, and the quoted GPT-6 Astra per-task scores swing widely, so the 20-27 point category margins and the 16.96 point Overall lead carry no error bars. The six waived tasks are assigned zero, which can only lower RoboICL's means, so that part of the protocol is conservative, not favorable. Second, selection on the evaluation set: J=B=12, B=25, and the anchored-versus-recent memory choice were all decided on four tasks that reappear in Table 2's category means and Overall. This does not disqualify the method, but the reported margins should be read as unconfirmed until a held-out allocation check and multi-seed rerun land. Third, the 20-27 point improvement over official zero-shot Astra is a whole-system comparison (Figure 3: different image packaging, action parameterization, execution rate, horizons), so the abstract's attribution to in-context organization is composite. Disclosed and legitimate as a framework claim; just not an ablation of ICL.
Who this is for: anyone working on frozen-VLM control or context engineering for embodied agents; the appendices are worth reading for the serialization details. The method is real, the caveats are stated, and the fix is cheap.
Recommendation: yes, send to peer review. Referees should require the multi-seed rerun and the held-out allocation check before accept, but this deserves referee time, not a desk rejection.
The strongest objection
The paper's central claim — RoboICL beats official zero-shot GPT-6 Astra by 20–27 progress points in every category and leads the leaderboard Overall (50.64 vs 33.68) — depends on RoboICL's scores being representative. Table 2 states 'All RoboICL evaluations use seed 0', baselines use three seeds and 50 episodes, and RoboICL runs 50 or 20 evaluations depending on task stability, with six tasks waived to zero. The framework's free parameters were selected on four development tasks — Build Tower, Cover Blocks, Put Bottles into Dustbin, Insert Tubes (Figure 4, Section 4.1) — that reappear in the headline Table 2 task set (Precision and Long-Horizon category means, Memory's Cover Blocks, and the Overall). Neither seed-0 representativeness nor the out-of-sample behavior of the dev-selected allocations is quantified; the authors state this directly (b149: 'validation on held-out tasks and layouts remains necessary to establish generality'). GPT-6 Astra per-task scores are visibly high-variance (Build Tower 16 to 100 and Classify Objects 100 to 71 within the ten-task panel), so a single-seed mean carries no error bar, and a favorable seed combined with test-informed parameter selection could plausibly inflate the reported margins. The 20–27-point comparison to official zero-shot Astra is also a system-level comparison (Figure 3: different action parameterization, adaptive horizons, execution at 25 Hz), so it does not isolate the contribution of the in-context organization — disclosed and legitimate as a framework claim, but the attribution asserted in the abstract is a composite of ICL and controller redesign. Because the margins are large, the waivers are conservative (they can only lower RoboICL's means), and the efficiency results carry independent support (Appendix C.1 verifies byte-identical cache frontiers across 1,780 requests), this is a CONDITIONAL-level representativeness concern, not a rejection of the method.
I read the paper as an empirical engineering claim: a frozen GPT-6 Astra, organized through demonstration context and bounded anchored interaction memory, produces large RoboDojo gains over the official zero-shot controller and the strongest leaderboard baselines without parameter updates. I examined each component of that claim for a load-bearing weakness. Baseline scores are quoted from the published RoboDojo leaderboard, the ten-task hybrid comparison is cited to Su et al. (2026), the waived-zero tasks are conservative (they can only lower RoboICL's means relative to evaluated baselines), the efficiency and cache claims carry independent support from Appendix C.1's byte-identical prefix checks across 1,780 consecutive requests, and the real-robot and Jev studies are explicitly scoped with their limits stated (b227, b259). The one place the central claim is not secure is representativeness: the headline numbers come from seed-0 runs with no variance estimate, at evaluation counts that differ from the comparison baselines, and with the context allocation and memory design chosen on four development tasks that then contribute to the Precision, Memory, and Long-Horizon category means and to the Overall. The authors flag the held-out-validation gap themselves (b149), so I am not raising a hidden flaw; I am confirming that the reader's weakest assumption is indeed the binding one. Because the claimed margins over the leaderboard are large (16.96 Overall points) and the waivers err conservative, even a substantial seed effect or mild tuning overfit is unlikely to flip the qualitative story, so I do not escalate to REJECT and I do not downgrade the reader's CONDITIONAL verdict. The decisive unresolved step is an independent multi-seed rerun of the 30-task panel with fixed hyperparameters, together with a check of whether the dev-selected allocations transfer to held-out seeds.
The referee's report
If the reported margins hold, this is a significant result: a purely inference-time, frozen-VLM controller that matches or beats robot-trained VLA/WAM baselines on a standard benchmark (Memory 70.00 vs 47.74, Open 54.75 vs 9.12, Precision comparable to Liber-0) would be a strong data point for context engineering over robot-specific post-training. The paper's methodology is unusually auditable: Appendix A.1 gives deterministic window-selection rules, Appendix C.1 verifies byte-identical cache prefixes across 1,780 consecutive requests, Appendix C.2 records every Jev-panel layout outcome, and code is released. I checked the category means and the 30-task Overall arithmetic against Table 2 and they are internally consistent. The main risk is statistical rather than algebraic: all RoboICL scores come from a single seed with no error bars, and the main configuration was chosen on four development tasks that reappear in the headline Table 2 benchmark. These two issues are load-bearing for the headline margins and need to be addressed.
Pith's reading of the paper (a machine's summary, not the authors')
The paper argues that the gap between general-purpose vision-language models and trained robot policies on precision and long-horizon tasks can be narrowed with no training at all, purely through how inference-time context is organized. RoboICL feeds a frozen GPT-6 Astra a shared observation–action–receipt–observation grammar applied to two sources: recorded demonstrations and the model's own accumulated rollout experience, kept in a bounded set of temporal anchors plus the latest interaction. On 30 RoboDojo tasks it reports improving on official zero-shot GPT-6 Astra by 20–27 progress-score points in every category and reaching a 30-task Overall score of 50.64 against a strongest baseline of 33.68, without a learned vision-language-action model or robot-specific parameter updates. If true, the claim matters because it would mean a general frontier model can handle high-precision and long-horizon manipulation through context design alone, and the same recipe could transfer across embodiments and models.
What carries the argument. The load-bearing object is the shared interaction grammar combined with bounded anchored memory. Every retained record, whether from a demonstration (synthetic receipt) or from live rollout (measured execution receipt), has the same before–action–after shape, so the model reads procedural examples and its own errors through one interface. Memory is bounded by B slots: the first B−1 are fixed temporal anchors distributed across the episode at targets τ_i = ⌊iT/B⌋, and the last slot holds the latest completed interaction for immediate correction, with <TRAJECTORY_GAP> tokens marking discarded spans. This keeps early context usable in long-horizon episodes while staying inside the API's 50-image limit, and the request layout (stable demonstration prefix, mutable memory suffix) is arranged so completed context stays byte-identical for KV-cache reuse.
Core claim. The central claim is that a frozen general-purpose VLM, given no parameter updates and no learned action head, can close most of its known deficit on precision and long-horizon robot tasks when its inference-time context is structured as a robot-native interface. RoboICL splits context into demonstration context (recorded examples, fixed per episode) and interaction memory (the model's own actions with their realized outcomes, updated after each executed chunk), and serializes both through one interaction grammar c_j = ((o_j,z_j), A_j, r_j, (o_{j+1},z_{j+1})): an observation and proprioceptive state, a predicted action chunk, an execution receipt reporting what actually ran, and the resulting observation. Bounded anchored memory retains the first B−1 interactions at fixed temporal targets τ_i = ⌊iT/B⌋ plus the latest interaction, with explicit <TRAJECTORY_GAP> records so omitted intervals are not misread as adjacent. On this design the paper reports a 50.64 Overall score across 30 RoboDojo tasks against 33.68 for the strongest compared baseline, leading on Memory and Open, matching the strongest Precision method, and gaining 20–27 points over official zero-shot GPT-6 Astra in every category; on three real-robot tasks mean progress rises from 14.45 at zero shot to 63.33 at one shot and 78.89 at three shots.
The premise it rests on. The headline benchmark scores come from single-seed (seed 0) RoboICL runs computed at variable per-task evaluation counts with hyperparameters tuned on four development tasks that also appear in the reported benchmark, compared against baselines run at three seeds and 50 episodes; the authors explicitly acknowledge that validation on held-out tasks and layouts is still needed.
If the paper is right (Pith's inferences, not the authors')
If the reported margins hold, context organization becomes a controllable and model-agnostic lever for embodied control, applicable to any frozen VLM without retraining.
Bounded anchored memory with gap markers offers a concrete recipe for long-horizon VLM control that does not grow with episode length, suggesting a path past the context-window ceiling.
The shared demonstration–interaction grammar implies that procedures and online error correction can be learned from context in the same format, which may reduce the need for robot-specific post-training across tasks.
The zero-shot result on Open, where interaction memory alone is used, indicates that closed-loop feedback can partially substitute for demonstrations in open-ended tasks.
Demonstration scaling on a real robot, with progress rising from 14.45 to 78.89 across zero to three shots, suggests the approach transfers across embodiments with only new demonstrations.
The reported gain over official zero-shot GPT-6 Astra is not isolated to memory design: the comparison bundles interface choices (stitched triptych packing, per-step Cartesian-increment action chunks, adaptive horizon) with the anchored-memory scheme, and the controlled ablation isolates only anchored versus recent retention. A like-for-like test varying one interface factor at a time would locate
The two-model Jev experiment hints at a general economy pattern: a cheap decision gate can halve calls to an expensive VLM while keeping or improving success on some tasks, a division of labor that plausibly extends beyond the two development tasks evaluated.
If the margins survive multi-seed held-out evaluation, the same bounded-anchored context recipe could be transferred to other frozen VLMs and other high-precision domains beyond RoboDojo, since nothing in the mechanism is benchmark-specific.
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
General-purpose vision-language models offer a promising way to zero-shot robot control: \gptastra excels at open-ended and language- or image-conditioned manipulation but remains substantially weaker on high-precision and long-horizon tasks. We introduce RoboICL, an in-context robot-control framework that narrows these gaps without robot-specific parameter updates or a learned VLA. RoboICL separates demonstration context, which provides recorded examples when available, from interaction memory, which accumulates the model's own actions and observed outcomes. Both use a shared observation–action–receipt–observation grammar. To preserve experience across task stages, RoboICL combines sampled demonstration blocks with bounded anchored memory. Fixed anchors keep earlier rollout interactions available for in-context learning, while the latest interaction supports immediate error correction. Across 30 RoboDojo tasks, using zero shot for Open and one demonstration elsewhere, RoboICL improves on official zero-shot \gptastra by 20–27 progress-score points in every category. It leads the leaderboard baselines on Memory and Open, achieves comparable performance to the strongest Precision baseline, and remains competitive on Long-Horizon. Its 30-task Overall score is 50.64, versus 33.68 for the strongest baseline. On a separate ten-task subset, RoboICL scores 60.60, within 2.00 points of the π0.5 + \gptastra hybrid approach. On three real-robot tasks, mean progress rises from 14.45 at zero shot to 63.33 at one shot and 78.89 at three shots. On two development tasks, optional Jev-gated action reuse reduces \gptastra calls by 33–48%. Code is available at \hrefhttps://github.com/Mosi-AI/RoboICLhttps://github.com/Mosi-AI/RoboICL.
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 · 2026-09-29
The authors' main claim, as the classifier picked it out 0.45
“Across 30 RoboDojo tasks, using zero shot for Open and one demonstration elsewhere, RoboICL improves on official zero-shot \gptastra{} by 20--27 progress-score points in every category.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.28 · human participants 0.05 · controlled comparison 0.97 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.99
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 |
|---|---|---|---|---|
| General-purpose vision-language models offer a promising way to zero-shot robot control: \gptastra{} excels at open-ended and language- or image-conditioned manipulation but remains substantially weaker on high-precision and long-horizon tasks. | background 0.87 | 0.19 | 0.05 | asserted 0.75 |
| We introduce \emph{RoboICL}, an in-context robot-control framework that narrows these gaps without robot-specific parameter updates or a learned VLA. | contribution 1.00 | 0.85 | 0.22 | asserted 1.00 |
| RoboICL separates \emph{demonstration context}, which provides recorded examples when available, from \emph{interaction memory}, which accumulates the model's own actions and observed outcomes. | method 0.97 | 0.80 | 0.05 | asserted 1.00 |
| Both use a shared observation--action--receipt--observation grammar. | method 1.00 | 0.55 | 0.02 | asserted 1.00 |
| To preserve experience across task stages, RoboICL combines sampled demonstration blocks with bounded anchored memory. | method 0.99 | 0.80 | 0.05 | asserted 1.00 |
| Fixed anchors keep earlier rollout interactions available for in-context learning, while the latest interaction supports immediate error correction. | method 1.00 | 0.64 | 0.04 | asserted 1.00 |
| Across 30 RoboDojo tasks, using zero shot for Open and one demonstration elsewhere, RoboICL improves on official zero-shot \gptastra{} by 20--27 progress-score points in every category. | result 1.00 | 0.96 | 0.39 | asserted 1.00 |
| It leads the leaderboard baselines on Memory and Open, achieves comparable performance to the strongest Precision baseline, and remains competitive on Long-Horizon. | result 0.99 | 0.91 | 0.26 | asserted 1.00 |
| Its 30-task Overall score is 50.64, versus 33.68 for the strongest baseline. | result 1.00 | 0.94 | 0.20 | asserted 1.00 |
| On a separate ten-task subset, RoboICL scores 60.60, within 2.00 points of the $\pi_{0.5}$ + \gptastra{} hybrid approach. | result 1.00 | 0.94 | 0.08 | asserted 1.00 |
| On three real-robot tasks, mean progress rises from 14.45 at zero shot to 63.33 at one shot and 78.89 at three shots. | result 1.00 | 0.96 | 0.11 | asserted 1.00 |
| On two development tasks, optional Jev-gated action reuse reduces \gptastra{} calls by 33--48\%. | result 0.95 | 0.94 | 0.03 | asserted 0.98 |
| Code is available at \href{https://github.com/Mosi-AI/RoboICL}{https://github.com/Mosi-AI/RoboICL}. | contribution 0.65 | 0.51 | 0.02 | asserted 0.89 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.95
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, 2026-09-29: 1 request, 9,706 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 (11)
Figures are the authors', reproduced from the paper.
References (16)
- [3] URLhttps://arxiv.org/abs/2204.14198. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agar- wal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Chil... on PithPith's reviewarXiv
- [9] URLhttps://arxiv.org/ abs/2606.25215. Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as Policies: Language Model Programs for Embodied Control.arXiv preprint arXiv:2209.07753, on PithPith's reviewarXiv
- [10] URLhttps://arxiv.org/abs/2209.07753. Yicheng Liu, Zibin Dong, Baijun Ye, Tianyuan Yuan, Tao Jiang, Anqi Yang, Shicheng Cao, Haonan Liu, Yue Sun, Zihan Guo, et al. G0. 5: One autoregressive stream for robot reasoning and action. arXiv preprint arXiv:2608.11739, on PithPith's reviewarXiv
- [15] URLhttps://arxiv.org/abs/2210.03629. Yida Yin, Zekai Wang, Yuvan Sharma, Dantong Niu, Trevor Darrell, and Roei Herzig. In-Context Learning Enables Robot Action Prediction in LLMs. InIEEE International Conference on Robotics and Automation, on PithPith's reviewarXiv
- [16] URLhttps://arxiv.org/abs/2410.12782. Wenbo Zhang, Kaixuan Wang, Yutao Ouyang, Xiaoyu Huang, Liyang Li, Kailun Su, Weiyang Jin, Wenhao Chai, Haotian Liang, Zhiyang Dou, Yue Chen, and Tianxing Chen. An unex- pected robot policy: Early evaluations of GPT-6 Astra on RoboDojo and b... on PithPith's reviewarXiv
- [18] URLhttps://arxiv.org/abs/2608.26103. 16 AI Model TFSeptember 2026 A METHOD AND IMPLEMENTATION A.1 CONTEXT CONSTRUCTION Interaction memory.Interaction memory exposes how earlier decisions change the environment. After decision roundt, the execution system returns receiptrt and ... on PithPith's reviewarXiv
- [25] receive 10 at the 1050-step limit. All panels use native head-camera frames. Contact strategy in Fasten Screws.The selected layout-2 pair contrasts zero-shot control, which scores 20, with a three-shot rollout that scores 100 (Figure 10, top). The three-shot policy brings the ... on Pith
- [100] Middle: Fill Pen Holder demonstrations with opposite hand assignments. Bottom: the selected rollout before and after transferring the filled holder from the right hand to the left. Each observation concatenates the left-wrist, head, and right-wrist views in that order. yellow ... on Pith
- [2017] URL https://proceedings.neurips.cc/paper_files/paper/2017/hash/ ba3866600c3540f67c1e9575e213be0a-Abstract.html. Letian Fu, Huang Huang, Gaurav Datta, Lawrence Yunliang Chen, William Chung-Ho Panitch, Fangchen Liu, Hui Li, and Ken Goldberg. In-Context Imitation Learning via Nex... on PithPith's reviewarXiv
- [2020] URLhttps://proceedings.neurips.cc/paper_files/paper/2020/ hash/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html. Fu Chen, Xin Ding, Bingjia Huang, Xiangyu Li, Mingju Wang, Jiawei He, Kun Li, Wei Sun, Yunxin Liu, Hao Wu, and Ting Cao. Zeva: In-Context Causal Learning for Generali... on PithPith's reviewarXiv
- [2021] URL https://proceedings.neurips.cc/paper_files/paper/2021/hash/ 01b7575c38dac42f3cfb7d500438b875-Abstract.html. Vitalis V osylius and Edward Johns. Instant Policy: In-Context Imitation Learning via Graph Diffusion. InInternational Conference on Learning Representations, on Pith
- [2022] URLhttps://arxiv.org/abs/2204.01691. Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Maria... on PithPith's reviewarXiv
- [2023] URLhttps://arxiv.org/abs/2303.11366. Kaustubh Sridhar, Souradeep Dutta, Dinesh Jayaraman, and Insup Lee. RICL: Adding In-Context Adaptability to Pre-Trained Vision-Language-Action Models. InProceedings of the 9th Con- ference on Robot Learning, volume 305, pp. 5022–5038. PMLR, on PithPith's reviewarXiv
- [2024] URLhttps://arxiv. org/abs/2403.19578. 14 AI Model TFSeptember 2026 Yan Duan, Marcin Andrychowicz, Bradly C. Stadie, Jonathan Ho, Jonas Schneider, Ilya Sutskever, Pieter Abbeel, and Wojciech Zaremba. One-Shot Imitation Learn- ing. InAdvances in Neural Information Processing Sys... on PithPith's reviewarXiv
- [2025] URLhttps: //www.robot-learning.uk/instant-policy. 15 AI Model TFSeptember 2026 Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. InInternational Conference on Learning Represe... on Pith
- [2026] URLhttps://arxiv.org/abs/2609.19138. Dexmal Team. DM0.5: An open-world foundation model for general-purpose embodied intelli- gence, July on PithPith's reviewarXiv
Cited by 18 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- Training Reasoning Models in Simulation to Natively Use Real Robots cs.RO · Nov 2026
- Embodied First Exam cs.RO · Nov 2026
- Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents cs.AI · Nov 2026
- Open-HumanWAM: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models cs.CV · Nov 2026
- LogLook: Execution-Grounded Memory for Frontier-Model Robot Planners cs.RO · Oct 2026
- When to Ask for Help: A Decision-Theoretic Account and Measurement Standard for Slow-Module Scheduling in Dual-System Vision-Language-Action Models cs.RO · Oct 2026
- Review the Episodes, Refine the Lessons: Simple Memory Evolution for Embodied Agents cs.RO · Nov 2026
- Ego-CEVR: Learning Cross-Embodiment Visual Representations from Human Egocentric Videos cs.RO · Oct 2026
- SO-101 Bench: Accessible Spatial Manipulation cs.CV · Oct 2026
- Profiling a Frontier Agent with EBench: From Direct Control to Harness with Policy cs.RO · Oct 2026
- MOTIO: Agentic Skill Engineering for Reality-Grounded Humanoid Loco-Manipulation cs.RO · Oct 2026
- Can a Frontier VLM Serve as a Robot Policy for Memory-Dependent Manipulation? cs.RO · Oct 2026
- ReTrace: Cross-Stage Failure Attribution for Self-Improving Robotic Manipulation cs.RO · Oct 2026
- Neuro Inspired Robotics: What Robot Learning Can Borrow from the Sensorimotor Brain cs.RO · Nov 2026
- Large Multimodal Models for Direct Robot Control: GPT-6 as a Beginning cs.RO · Oct 2026
- Frontier LLMs Enter the World of Robotics: A Cross-Model Evaluation on RoboColiseum cs.RO · Oct 2026
- GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation cs.RO · Oct 2026
- One Policy for 100 Household Tasks: Our Solution for the 2026 BEHAVIOR Challenge cs.RO · Nov 2026
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