Pith.

Robotics · arXiv:2609.19554 v1 · submitted 17 Sep 2026 · 76 references

VABench: Measuring Embodied Spatial Intelligence through Visual Demonstrations, Active Perception, and Metric Control

Zhongbo Zhang, Jiayi Jin, Yifan Wang, Zaibin Zhang, Haiwen Diao, Lijun Wang, Huchuan Lu

The paper says

“Spatial intelligence requires more than describing object locations. Under incomplete observation, models must identify and acquire missing evidence, interpret it in a common spatial frame, and act on it. We introduce VA-Bench to evaluate the complete observe-reason-act-revise loop.”

more of the abstract

General-purpose MLLMs learn procedural context from RGB-only demonstrations, actively select camera viewpoints, issue metric Cartesian commands, and revise them from execution feedback. Models receive no privileged object poses, oracle trajectories, or learned action heads. A fixed model-agnostic controller executes only model-specified targets. VA-Bench contains 14 base task families (11 single-arm and three dual-arm), seven held-out geometry/layout variants, and a long-horizon five-object composition track. We evaluate 12 primary model conditions in three independent runs over the same 20 physically verified seeds per base task, reporting terminal success, nine trajectory-level behavioral diagnostics, and subtask progress. First, the best-performing model scores 100.0% on target localization and 78.9% on spatial relations in the annotated run. Its three-run macro-average task success is only 53.93+/-3.17%. Second, active camera control significantly improves task success over passive multi-view observation. In one matched comparison, success rises from 27.86% to 57.50%. Third, held-out geometric transfer can reduce task success by over 30 percentage points. No model completes a strict long-horizon episode, despite substantial partial progress. VA-Bench thus tests whether general-purpose MLLMs can turn visual demonstrations and actively acquired evidence into successful embodied action.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 18 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. Solid, gap-filling benchmark with an honest protocol; the real gap is that seed verification proves physical solvability, not RGB-only visual solvability, so the 53.93% headline is best read as an end-to-end interface-and-model result. This line is wrong

1 major objection · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 76 references say.

The paper says, at §3.6, Table 21

The review did not quote this passage.

Pith says: Objection 1 of 1 · major

The seed verification (§3.6) certifies physical solvability using an expert with privileged state (full object poses, exact coordinates). The paper does not verify that each retained seed is solvable under the MLLM's observation interface (RGB-only, no depth, no camera calibration, bounded gripper-centered viewpoints).

more of objection 1

If some instances require depth information, precise metric triangulation that the limited viewpoint set cannot support, or visual disambiguation infeasible from the provided RGB frames, then even an ideal RGB-only agent could not succeed. Failures on such instances would inflate the reported failure rate and partially undermine the claim that the gap between diagnostics and task execution reflects a deficit in spatial intelligence. The authors should either add a visual-solvability analysis (e.g., a human oracle or a perfect visual grounding planner on a subset of seeds) or discuss this limitation explicitly in §B.3, quantifying the potential impact on the headline success rates.

5 minor objections

The paper says, at Abstract and §1

Pith says: minor objection 1 of 5

The notation 'V A-Bench' appears with an extra space in several places (e.g., 'V A-Bench' in the abstract and §1). Ensure consistent formatting as 'VA-Bench' throughout.

The paper says, at §3.6

Pith says: minor objection 2 of 5

The seed acceptance rate of 88.1% is conditional on candidates that were probed, not on the full candidate pool. Clarify that unvisited candidates are not counted in the denominator (currently implied by 'conditional yield' but could be made more explicit in the main text).

show 3 more

The paper says, at Table 3

Pith says: minor objection 3 of 5

The table lists 'Doubao-2.1' and 'Gemini-3.6' with underscores; in the text they are written as 'Doubao-2.1' and 'Gemini-3.6'. Unify formatting for readability.

The paper says, at §B.1

Pith says: minor objection 4 of 5

The limitations paragraph states 'The same applies to the protocol comparisons.' but does not fully restate the open mechanism for the active–passive result. Consider adding a one-sentence summary of what the comparison does not isolate (e.g., view relevance, cross-view integration, context length).

The paper says, at Appendix C.9.1, Table 28

Pith says: minor objection 5 of 5

The Fleiss κ for TL is 0.723 with 97.41% unanimity and near-ceiling positive rate. This is noted, but the interpretation could be more precise: κ is depressed by prevalence, not by low agreement. The Gwet AC1 (0.982) is more appropriate. The text already acknowledges this, but it is worth re-emphasizing that the low κ does not indicate poor agreement.

What would settle it

Select 50 failed episodes from the top three conditions (Qwen3.8-max, Opus-5, GPT-5.6-sol) across different task families. For each episode, provide a human operator with the same demonstration summary, task instruction, observation interface (RGB-only, same camera action set, same proprioception), and action budget. Record whether the human can manually infer and dispatch the correct metric commands to achieve task success. If the human succeeds on ≥45/50, visual solvability is confirmed and the concern is mitigated; the MLLM failures then stem from limited spatial reasoning. If the human succeeds on <35/50, the benchmark contains instances that are not visually solvable, and the central claim about MLLM capability should be qualified accordingly.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 18 Sep 2026

  1. Reader holds up confidence high
  2. Skeptic partly agrees 48 s · 39,321 tokens read, 4,763 written
  3. Editor wrote the verdict above 37 s · 40,393 tokens read, 3,849 written
  4. Referee minor revision: 1 major objection, 5 minor 30 s · 39,744 tokens read, 3,014 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

VA-Bench is a genuine contribution. It is the first benchmark I know of that makes a general-purpose MLLM close the entire loop — extract a procedure from an RGB demonstration, actively choose viewpoints, emit metric Cartesian commands, and revise them from execution feedback — with no oracle poses, depth, or learned action head. The active-versus-passive comparison is carefully controlled: the passive five-view bundle is a strict subset of the views reachable in active mode, seeds, executor, and budget are matched, and the sign tests back the claim that model-directed viewing matters. Three-run reporting with sample dispersion is honest, and the authors correctly decline to order their top three models.

Several things are better than the usual benchmark paper. The behavioral diagnostics have strong inter-annotator agreement, and the appendix checks whether annotations merely restate terminal outcome — they do not: TL and MS are positive on over a thousand failing episodes. The cross-model summary-transfer result is genuinely interesting — the best-executing model's summary hurt every other reader — and is appropriately flagged as suggestive rather than definitive.

The stress-test concern is real and worth fixing, though it is a missing control, not a demonstrated flaw. Section 3.6 certifies that each retained seed is physically solvable by a privileged expert; it does not certify that the task-relevant geometry is recoverable from the RGB-only observation and action interface. If some seeds require estimates the gripper-centered viewpoint set cannot support, part of the reported failure rate is a property of the interface rather than of spatial understanding. The authors state "visually recoverable spatial bottleneck" as a design requirement but never verify it per seed. That is fixable — an RGB-only oracle perception module, or a per-family argument that the needed geometry is recoverable from the allowed views, would close the gap. The paper's own framing in B.1 and D.4 already hedges by calling outcomes an end-to-end system measure; the abstract's "embodied spatial intelligence" framing goes a bit further than the evidence supports. Minor: some ablations run once, and the long-horizon track is a single family; both are acknowledged.

This paper deserves a serious referee. It is careful, largely reproducible, and fills an actual gap in the evaluation landscape. I would ask the authors to add the visual-solvability check rather than reject; the central finding that the top model's 53.93% lags its own 100% localization and 78.9% spatial-relations diagnostics will survive it.

The strongest objection

The central claim that current MLLMs achieve at most 53.93% success and complete no long-horizon episodes is interpreted as a limitation on embodied spatial intelligence. However, the seed verification in §3.6 uses an expert with full state information (privileged object poses, exact coordinates) to certify physical solvability. The verification does not test whether the task instance is solvable under the actual observation interface the MLLM faces: RGB-only, no depth, no camera calibration, no object poses, and only a bounded set of gripper-centered semantic viewpoints. If some instances require depth information, precise metric triangulation that the limited viewpoint set cannot support, or visual disambiguation that is infeasible from the provided RGB frames, then no RGB-only agent—including an ideal one—could succeed. Failures on such instances would then reflect a benchmark design constraint rather than a deficit in MLLM spatial understanding. The paper acknowledges that the active–passive comparison leaves mechanism open (§B.1), but does not verify that every retained seed is spatially solvable under the visual constraints imposed on the MLLM. Without this verification, a substantial fraction of the reported failure rate could be attributed to visual ambiguity rather than to failures in reasoning, planning, or control. This directly threatens the headline claim about the gap between spatial diagnostics and task execution.

The reader identified physical solvability as the weakest assumption, focusing on whether simulation constraints could make seeds unsolvable for an ideal agent. My concern extends this: even if physically solvable, seeds may not be solvable under the MLLM's specific observation constraints (RGB-only, no depth, no calibration, limited views). This is a more specific and load-bearing threat to the central claim because the paper interprets failure as evidence of limited embodied spatial intelligence. However, the paper is otherwise well-designed, transparent about limitations, and provides extensive behavioral diagnostics. The proposed concrete test would settle whether this concern actually lands. If the test shows most seeds are visually solvable, the paper's claims stand. If not, the interpretation would need adjustment, but the benchmark data itself remains valuable. I therefore recommend the verdict remain UNCHANGED; this is a suggestion for an additional control experiment, not grounds for rejection.

The referee's report

If the results hold, VA-Bench provides a well-designed, reproducible evaluation protocol that measures how MLLMs integrate active perception, metric reasoning, and execution feedback. The finding that strong diagnostic scores on spatial perception (100% target localization) coexist with weak task completion (≈54% macro-average) is informative and suggests that current models fail to close the loop between understanding and action. The active–passive comparison cleanly demonstrates the value of model-directed evidence acquisition. The benchmark is self-contained, uses physics-based success criteria, and has high inter-annotator agreement on behavioral diagnostics. Together, these features make VA-Bench a useful tool for the community to track progress on embodied spatial intelligence.

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

The paper introduces VA-Bench, a physics-simulated benchmark that requires multimodal large language models (MLLMs) to actively gather visual evidence, translate spatial understanding into metric robot commands, and revise actions from execution feedback. Across 14 task families and 12 model conditions, the best model achieves only 53.93% macro-average success, despite scoring 100% on target localization and nearly 80% on spatial relations in annotated runs. No model completes a strict long-horizon composition task. The authors show that active camera control substantially improves performance over passive multi-view input, and that held-out geometric variations can reduce success by over 30 percentage points. The benchmark thus isolates a critical gap: general-purpose MLLMs can describe the world but not reliably act on it through a closed loop.

What carries the argument. VA-Bench, a benchmark built on the RoboTwin simulator, comprising 14 base task families (11 single-arm, 3 dual-arm), held-out geometry variants, and a five-object long-horizon composition track. The evaluation protocol requires an MLLM to infer procedures from RGB-only demonstrations, actively select camera viewpoints, issue metric Cartesian translation and rotation commands, and revise plans based on execution outcomes (validity, planner feedback, visual changes). A fixed, model-agnostic controller executes only the model-specified targets. The benchmark reports terminal task success, nine behavioral diagnostics (target localization, informative active exploration, robot–object spatial relations, manipulation semantics, goal-directed planning, fine-grained pre-contact analysis, error detection, online correction, post-failure adjustment), and subtask progress.

Core claim. The central discovery is that current general-purpose MLLMs, when evaluated in a closed-loop observe-reason-act-revise protocol without privileged object poses or learned action heads, fail to translate strong spatial perception and manipulation semantics into reliable embodied task execution. The best models succeed on about half of the base tasks, and their performance degrades sharply under geometric transfer and long-horizon demands. Active evidence acquisition (camera control) is critical; removing it reduces success by 10–30 percentage points across tested models. The trajectory-level diagnostics reveal that the bottleneck lies not in target finding but in robot–object spatial relations, fine-grained pre-contact analysis, and online correction.

The premise it rests on. Every seed in the base suite is verified to be physically solvable by an ideal agent; if some seeds are unsolvable due to simulation errors not caught by the verification pipeline, then model failure rates could be overestimated.

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

General-purpose MLLMs in their current form are not reliable for independent manipulation; their strong visual-semantic scores do not guarantee precise metric control and recovery.

Active camera control provides a substantial advantage over passive multi-view observation, suggesting that the ability to direct attention is a separable and necessary capability for embodied spatial intelligence.

Geometric and layout variations cause large drops in success, indicating that current models fail to re-ground procedural knowledge in unfamiliar spatial configurations.

Long-horizon composition remains unsolved: even the best models place only 61% of objects correctly and complete zero full episodes, pointing to a failure in maintaining state and sequencing skills.

The benchmark's trajectory-level diagnostics reveal that the bottleneck lies not in target finding but in robot–object spatial relations, pre-contact analysis, and online correction.

If the bottleneck is in metric grounding and recovery rather than visual semantics, then future progress may depend on architectural changes (e.g., explicit 3D geometric reasoning, learned control policies) rather than scaling existing MLLMs.

The cross-model summary transfer experiment showed that a weaker model's summary can improve other models while a stronger model's summary hurts them, suggesting that summary explicitness matters more than author capability — a finding that could inform shared procedural interface design.

The active–passive gap points toward training MLLMs on interactive data where they must decide when to look, not just what to say.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Spatial intelligence requires more than describing object locations. Under incomplete observation, models must identify and acquire missing evidence, interpret it in a common spatial frame, and act on it. We introduce VA-Bench to evaluate the complete observe-reason-act-revise loop. General-purpose MLLMs learn procedural context from RGB-only demonstrations, actively select camera viewpoints, issue metric Cartesian commands, and revise them from execution feedback. Models receive no privileged object poses, oracle trajectories, or learned action heads. A fixed model-agnostic controller executes only model-specified targets. VA-Bench contains 14 base task families (11 single-arm and three dual-arm), seven held-out geometry/layout variants, and a long-horizon five-object composition track. We evaluate 12 primary model conditions in three independent runs over the same 20 physically verified seeds per base task, reporting terminal success, nine trajectory-level behavioral diagnostics, and subtask progress. First, the best-performing model scores 100.0% on target localization and 78.9% on spatial relations in the annotated run. Its three-run macro-average task success is only 53.93+/-3.17%. Second, active camera control significantly improves task success over passive multi-view observation. In one matched comparison, success rises from 27.86% to 57.50%. Third, held-out geometric transfer can reduce task success by over 30 percentage points. No model completes a strict long-horizon episode, despite substantial partial progress. VA-Bench thus tests whether general-purpose MLLMs can turn visual demonstrations and actively acquired evidence into successful embodied action.

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 · 2026-09-23

The authors' main claim, as the classifier picked it out 0.49

“VA-Bench thus tests whether general-purpose MLLMs can turn visual demonstrations and actively acquired evidence into successful embodied action.”

How the main claim is stated: asserted 0.98

What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.83 · human participants 0.05 · controlled comparison 0.94 · randomized assignment 0.03 · states quantitative result 0.99 · compares with prior results 0.84 · mentions code or data availability 0.05

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Spatial intelligence requires more than describing object locations. background 0.92 0.06 0.03 asserted 1.00
Under incomplete observation, models must identify and acquire missing evidence, interpret it in a common spatial frame, and act on it. background 0.96 0.07 0.04 asserted 0.98
We introduce VA-Bench to evaluate the complete observe-reason-act-revise loop. contribution 1.00 0.91 0.06 asserted 0.99
General-purpose MLLMs learn procedural context from RGB-only demonstrations, actively select camera viewpoints, issue metric Cartesian commands, and revise them from execution feedback. method 0.95 0.65 0.06 asserted 1.00
Models receive no privileged object poses, oracle trajectories, or learned action heads. method 0.99 0.60 0.03 asserted 1.00
A fixed model-agnostic controller executes only model-specified targets. method 0.99 0.54 0.02 asserted 1.00
VA-Bench contains 14 base task families (11 single-arm and three dual-arm), seven held-out geometry/layout variants, and a long-horizon five-object composition track. method 0.81 0.81 0.02 asserted 1.00
We evaluate 12 primary model conditions in three independent runs over the same 20 physically verified seeds per base task, reporting terminal success, nine trajectory-level behavioral diagnostics, and subtask progress. method 0.97 0.96 0.03 asserted 0.99
First, the best-performing model scores 100.0% on target localization and 78.9% on spatial relations in the annotated run. result 1.00 0.93 0.06 asserted 1.00
Its three-run macro-average task success is only 53.93+/-3.17%. result 1.00 0.95 0.10 asserted 1.00
Second, active camera control significantly improves task success over passive multi-view observation. result 0.99 0.93 0.14 asserted 1.00
In one matched comparison, success rises from 27.86% to 57.50%. result 1.00 0.94 0.06 asserted 1.00
Third, held-out geometric transfer can reduce task success by over 30 percentage points. result 1.00 0.93 0.07 asserted 0.91
No model completes a strict long-horizon episode, despite substantial partial progress. result 0.99 0.93 0.19 asserted 1.00
VA-Bench thus tests whether general-purpose MLLMs can turn visual demonstrations and actively acquired evidence into successful embodied action. interpretation 0.60 0.51 0.42 asserted 0.98

What kind of paper the classifier reports this is: dataset or benchmark 1.00 · evidence from observation or experiment 0.99

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, 2026-09-23: 1 request, 10,604 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)
Figure 1 of the paper
Figure 1. Overview of the evaluation target. Fixed-view spatial evaluations can stop at a verbal
Figure 2 of the paper
Figure 2. Execution success, subtask progress, and behavioral diagnostics. (a) Mean terminal suc
Figure 3 of the paper
Figure 3. Annotation reference for the nine trajectory-level capability criteria. In the eight trajectory
Figure 4 of the paper
Figure 4. Schematic of the shared agent harness. Structured model decisions are validated and

Figures are the authors', reproduced from the paper.

References (76)
  1. [1] Zhao, Enyu and Raval, Vedant and Zhang, Hejia and Mao, Jiageng and Shangguan, Zeyu and Nikolaidis, Stefanos and Wang, Yue and Seita, Daniel , booktitle=. 2025 , url= on Pith
  2. [2] Yang, Rui and Chen, Hanyang and Zhang, Junyu and Zhao, Mark and Qian, Cheng and Wang, Kangrui and Wang, Qineng and Koripella, Teja Venkat and Movahedi, Marziyeh and Li, Manling and Ji, Heng and Zhang, Huan and Zhang, Tong , booktitle=. 2025 , url= on Pith
  3. [3] Zhang, Pingyue and Huang, Zihan and Wang, Yue and Zhang, Jieyu and Xue, Letian and Wang, Zihan and Wang, Qineng and Chandrasegaran, Keshigeyan and Zhang, Ruohan and Choi, Yejin and Krishna, Ranjay and Wu, Jiajun and Fei-Fei, Li and Li, Manling , journal= on Pith
  4. [4] Hong, Yining and Liu, Jiageng and Yin, Han and Li, Manling and Guibas, Leonidas and Fei-Fei, Li and Wu, Jiajun and Choi, Yejin , journal= on Pith
  5. [5] Gao, Hongcheng and Qu, Hailong and Tang, Jingyi and Wang, Jiahao and Huang, Zihao and others , journal= on Pith
  6. [6] Wu, Xin and Liang, Zhixuan and Ma, Yue and Hu, Mengkang and Qin, Zhiyuan and Li, Xiu , journal= on Pith
  7. [7] Maurya, Anurag and Jain, Sukhvansh and Avhad, Prajwal and Balachandran, Gautham and Zhou, Ziyi and Kshirsagar, Atharva and Singh, Satyam and Li, Bowen and Mukund, Rishabh and Singh, Ritul and Vira, Jatin and Chatterjee, Suvonil and Jha, Devesh K. , journal= on Pith
  8. [8] Zhao, Enyu and Raval, Vedant and Zhang, Hejia and Mao, Jiageng and Shangguan, Zeyu and Nikolaidis, Stefanos and Wang, Yue and Seita, Daniel , booktitle =. 2025 , url = on Pith
  9. [9] Yang, Rui and Chen, Hanyang and Zhang, Junyu and Zhao, Mark and Qian, Cheng and Wang, Kangrui and Wang, Qineng and Koripella, Teja Venkat and Movahedi, Marziyeh and Li, Manling and others , booktitle =. 2025 , url = on Pith
  10. [10] Zhang, Pingyue and Huang, Zihan and Wang, Yue and Zhang, Jieyu and Xue, Letian and Wang, Zihan and Wang, Qineng and Chandrasegaran, Keshigeyan and Zhang, Ruohan and Choi, Yejin and others , booktitle =. 2026 , url = on Pith
  11. [11] Hong, Yining and Liu, Jiageng and Yin, Han and Li, Manling and Guibas, Leonidas and Li, Fei-Fei and Wu, Jiajun and Choi, Yejin , journal =. 2026 , url = on Pith
  12. [12] Gao, Hongcheng and Qu, Hailong and Tang, Jingyi and Wang, Jiahao and Huang, Zihao and Qiao, Hengkang and Huang, Shihong and Yang, Junming and Li, Yi and Yuan, Hongyixuan and others , journal =. 2026 , url = on Pith
  13. [13] Wu, Xin and Liang, Zhixuan and Ma, Yue and Hu, Mengkang and Qin, Zhiyuan and Li, Xiu , journal =. 2026 , url = on Pith
  14. [14] Maurya, Anurag and Jain, Sukhvansh and Avhad, Prajwal and Balachandran, Gautham and Zhou, Ziyi and Kshirsagar, Atharva and Singh, Satyam and Li, Bowen and Mukund, Rishabh and Singh, Ritul and others , journal =. 2026 , url = on Pith
  15. [15] Visual Spatial Reasoning , author =. Transactions of the Association for Computational Linguistics , volume =. 2023 , doi = on Pith
  16. [16] What's ``Up'' with Vision-Language Models? Investigating Their Struggle with Spatial Reasoning , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages =. 2023 , doi = on Pith
  17. [17] Du, Mengfei and Wu, Binhao and Li, Zejun and Huang, Xuanjing and Wei, Zhongyu , booktitle =. 2024 , doi = on Pith
  18. [18] Song, Chan Hee and Blukis, Valts and Tremblay, Jonathan and Tyree, Stephen and Su, Yu and Birchfield, Stan , booktitle =. 2025 , url = on Pith
  19. [19] Fu, Xingyu and Hu, Yushi and Li, Bangzheng and Feng, Yu and Wang, Haoyu and Lin, Xudong and Roth, Dan and Smith, Noah A. and Ma, Wei-Chiu and Krishna, Ranjay , booktitle =. 2024 , doi = on Pith
  20. [20] Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =. 2025 , url = on Pith
show 56 more
  1. [21] Ma, Wufei and Chen, Haoyu and Zhang, Guofeng and Chou, Yu-Cheng and Chen, Jieneng and de Melo, Celso and Yuille, Alan , booktitle =. 2025 , url = on Pith
  2. [22] Azuma, Daichi and Miyanishi, Taiki and Kurita, Shuhei and Kawanabe, Motoaki , booktitle =. 2022 , url = on Pith
  3. [23] Ma, Xiaojian and Yong, Silong and Zheng, Zilong and Li, Qing and Liang, Yitao and Zhu, Song-Chun and Huang, Siyuan , booktitle =. 2023 , url = on Pith
  4. [24] Chen, Boyuan and Xu, Zhuo and Kirmani, Sean and Ichter, Brian and Sadigh, Dorsa and Guibas, Leonidas and Xia, Fei , booktitle =. 2024 , url = on Pith
  5. [25] Cheng, An-Chieh and Yin, Hongxu and Fu, Yang and Guo, Qiushan and Yang, Ruihan and Kautz, Jan and Wang, Xiaolong and Liu, Sifei , booktitle =. 2024 , doi = on Pith
  6. [26] Reasoning Paths with Reference Objects Elicit Quantitative Spatial Reasoning in Large Vision-Language Models , author =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages =. 2024 , doi = on Pith
  7. [27] Zheng, Kaizhi and Chen, Xiaotong and Jenkins, Odest Chadwicke and Wang, Xin , booktitle =. 2022 , doi = on Pith
  8. [28] Mees, Oier and Hermann, Lukas and Rosete-Beas, Erick and Burgard, Wolfram , journal =. 2022 , doi = on Pith
  9. [29] Gu, Jiayuan and Xiang, Fanbo and Li, Xuanlin and Ling, Zhan and Liu, Xiqiang and Mu, Tongzhou and Tang, Yihe and Tao, Stone and Wei, Xinyue and Yao, Yunchao and others , booktitle =. 2023 , url = on Pith
  10. [30] Mu, Yao and Chen, Tianxing and Chen, Zanxin and Peng, Shijia and Lan, Zhiqian and Gao, Zeyu and Liang, Zhixuan and Yu, Qiaojun and Zou, Yude and Xu, Mingkun and others , booktitle =. 2025 , url = on Pith
  11. [31] Zhang, Shiduo and Xu, Zhe and Liu, Peiju and Yu, Xiaopeng and Li, Yuan and Gao, Qinghui and Fei, Zhaoye and Yin, Zhangyue and Wu, Zuxuan and Jiang, Yu-Gang and others , booktitle =. 2025 , url = on Pith
  12. [32] Ichter, Brian and Brohan, Anthony and Chebotar, Yevgen and Finn, Chelsea and Hausman, Karol and Herzog, Alexander and Ho, Daniel and Ibarz, Julian and Irpan, Alex and Jang, Eric and others , booktitle =. Do As. 2023 , url = on Pith
  13. [33] Inner Monologue: Embodied Reasoning through Planning with Language Models , author =. Proceedings of the 6th Conference on Robot Learning , volume =. 2023 , url = on Pith
  14. [34] Code as Policies: Language Model Programs for Embodied Control , author =. 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages =. 2023 , doi = on Pith
  15. [35] Driess, Danny and Xia, Fei and Sajjadi, Mehdi S. M. and Lynch, Corey and Chowdhery, Aakanksha and Ichter, Brian and Wahid, Ayzaan and Tompson, Jonathan and Vuong, Quan and Yu, Tianhe and others , booktitle =. 2023 , url = on Pith
  16. [36] Rana, Krishan and Haviland, Jesse and Garg, Sourav and Abou-Chakra, Jad and Reid, Ian and S. Proceedings of the 7th Conference on Robot Learning , volume =. 2023 , url = on Pith
  17. [37] Huang, Wenlong and Wang, Chen and Zhang, Ruohan and Li, Yunzhu and Wu, Jiajun and Li, Fei-Fei , booktitle =. 2023 , url = on Pith
  18. [38] Nasiriany, Soroush and Xia, Fei and Yu, Wenhao and Xiao, Ted and Liang, Jacky and Dasgupta, Ishita and Xie, Annie and Driess, Danny and Wahid, Ayzaan and Xu, Zhuo and others , booktitle =. 2024 , url = on Pith
  19. [39] Fang, Kuan and Liu, Fangchen and Abbeel, Pieter and Levine, Sergey , booktitle =. 2024 , doi = on Pith
  20. [40] Yuan, Wentao and Duan, Jiafei and Blukis, Valts and Pumacay, Wilbert and Krishna, Ranjay and Murali, Adithyavairavan and Mousavian, Arsalan and Fox, Dieter , booktitle =. 2025 , url = on Pith
  21. [41] Huang, Wenlong and Wang, Chen and Li, Yunzhu and Zhang, Ruohan and Li, Fei-Fei , booktitle =. 2025 , url = on Pith
  22. [42] Yu, Tianhe and Quillen, Deirdre and He, Zhanpeng and Julian, Ryan and Hausman, Karol and Finn, Chelsea and Levine, Sergey , booktitle =. 2020 , url = on Pith
  23. [43] James, Stephen and Ma, Zicong and Rovick Arrojo, David and Davison, Andrew J. , journal =. 2020 , doi = on Pith
  24. [44] Mu, Tongzhou and Ling, Zhan and Xiang, Fanbo and Yang, Derek and Li, Xuanlin and Tao, Stone and Huang, Zhiao and Jia, Zhiwei and Su, Hao , booktitle =. 2021 , url = on Pith
  25. [45] Li, Chengshu and Zhang, Ruohan and Wong, Josiah and Gokmen, Cem and Srivastava, Sanjana and Mart. Proceedings of the 6th Conference on Robot Learning , volume =. 2023 , url = on Pith
  26. [46] Nasiriany, Soroush and Maddukuri, Abhiram and Zhang, Lance and Parikh, Adeet and Lo, Aaron and Joshi, Abhishek and Mandlekar, Ajay and Zhu, Yuke , booktitle =. 2024 , doi = on Pith
  27. [47] Shridhar, Mohit and Manuelli, Lucas and Fox, Dieter , booktitle =. 2022 , url = on Pith
  28. [48] Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation , author =. Proceedings of the 6th Conference on Robot Learning , volume =. 2023 , url = on Pith
  29. [49] Brohan, Anthony and Brown, Noah and Carbajal, Justice and Chebotar, Yevgen and Dabis, Joseph and Finn, Chelsea and Gopalakrishnan, Keerthana and Hausman, Karol and Herzog, Alexander and Hsu, Jasmine and others , booktitle =. 2023 , doi = on Pith
  30. [50] Jiang, Yunfan and Gupta, Agrim and Zhang, Zichen and Wang, Guanzhi and Dou, Yongqiang and Chen, Yanjun and Li, Fei-Fei and Anandkumar, Anima and Zhu, Yuke and Fan, Linxi , booktitle =. 2023 , url = on Pith
  31. [51] Zitkovich, Brianna and Yu, Tianhe and Xu, Sichun and Xu, Peng and Xiao, Ted and Xia, Fei and Wu, Jialin and Wohlhart, Paul and Welker, Stefan and Wahid, Ayzaan and others , booktitle =. 2023 , url = on Pith
  32. [52] Goyal, Ankit and Xu, Jie and Guo, Yijie and Blukis, Valts and Chao, Yu-Wei and Fox, Dieter , booktitle =. 2023 , url = on Pith
  33. [53] Diffusion Policy: Visuomotor Policy Learning via Action Diffusion , author =. Proceedings of Robotics: Science and Systems , year =. doi:10.15607/RSS.2023.XIX.026 , url = on Pithdoi:10.15607/rss.2023.xix.026
  34. [54] O'Neill, Abby and Rehman, Abdul and Gupta, Abhinav and Maddukuri, Abhiram and Gupta, Abhishek and Padalkar, Abhishek and Lee, Abraham and Pooley, Acorn and Gupta, Agrim and Mandlekar, Ajay and others , booktitle =. Open. 2024 , doi = on Pith
  35. [55] Ghosh, Dibya and Walke, Homer Rich and Pertsch, Karl and Black, Kevin and Mees, Oier and Dasari, Sudeep and Hejna, Joey and Kreiman, Tobias and Xu, Charles and Luo, Jianlan and others , booktitle =. 2024 , doi = on Pith
  36. [56] Kim, Moo Jin and Pertsch, Karl and Karamcheti, Siddharth and Xiao, Ted and Balakrishna, Ashwin and Nair, Suraj and Rafailov, Rafael and Foster, Ethan P. and Sanketi, Pannag R. and Vuong, Quan and others , booktitle =. 2025 , url = on Pith
  37. [57] Vision-Language Foundation Models as Effective Robot Imitators , author =. International Conference on Learning Representations , year = on Pith
  38. [58] Jang, Eric and Irpan, Alex and Khansari, Mohi and Kappler, Daniel and Ebert, Frederik and Lynch, Corey and Levine, Sergey and Finn, Chelsea , booktitle =. 2022 , url = on Pith
  39. [59] Mandlekar, Ajay and Nasiriany, Soroush and Wen, Bowen and Akinola, Iretiayo and Narang, Yashraj and Fan, Linxi and Zhu, Yuke and Fox, Dieter , booktitle =. 2023 , url = on Pith
  40. [60] Wang, Chen and Fan, Linxi and Sun, Jiankai and Zhang, Ruohan and Li, Fei-Fei and Xu, Danfei and Zhu, Yuke and Anandkumar, Anima , booktitle =. 2023 , url = on Pith
  41. [61] What Matters in Learning from Offline Human Demonstrations for Robot Manipulation , author =. Proceedings of the 5th Conference on Robot Learning , volume =. 2022 , url = on Pith
  42. [62] Liu, Zeyi and Bahety, Arpit and Song, Shuran , booktitle =. 2023 , url = on Pith
  43. [63] Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners , author =. Proceedings of the 7th Conference on Robot Learning , volume =. 2023 , url = on Pith
  44. [64] Duan, Jiafei and Pumacay, Wilbert and Kumar, Nishanth and Wang, Yi Ru and Tian, Shulin and Yuan, Wentao and Krishna, Ranjay and Fox, Dieter and Mandlekar, Ajay and Guo, Yijie , booktitle =. 2025 , url = on Pith
  45. [65] Xiong, Chuyan and Shen, Chengyu and Li, Xiaoqi and Zhou, Kaichen and Liu, Jiaming and Wang, Ruiping and Dong, Hao , booktitle =. Autonomous Interactive Correction. 2025 , url = on Pith
  46. [66] Jiang, Yunfan and Wang, Chen and Zhang, Ruohan and Wu, Jiajun and Li, Fei-Fei , booktitle =. 2025 , url = on Pith
  47. [67] Measuring Nominal Scale Agreement among Many Raters , author =. Psychological Bulletin , volume =. 1971 , doi = on Pith
  48. [68] Computing Inter-Rater Reliability and Its Variance in the Presence of High Agreement , author =. British Journal of Mathematical and Statistical Psychology , volume =. 2008 , doi = on Pith
  49. [69] Johnson, Justin and Hariharan, Bharath and van der Maaten, Laurens and Li, Fei-Fei and Zitnick, C. Lawrence and Girshick, Ross , booktitle =. 2017 , doi = on Pith
  50. [70] Hudson, Drew A. and Manning, Christopher D. , booktitle =. 2019 , doi = on Pith
  51. [71] Liu, Bo and Zhu, Yifeng and Gao, Chongkai and Feng, Yihao and Liu, Qiang and Zhu, Yuke and Stone, Peter , booktitle =. 2023 , doi = on Pith
  52. [72] Active Perception , author =. Proceedings of the IEEE , volume =. 1988 , doi = on Pith
  53. [73] A Solution to the Next Best View Problem for Automated Surface Acquisition , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 1999 , doi = on Pith
  54. [74] Tong, Shengbang and Brown, Ellis and Wu, Penghao and Woo, Sanghyun and Middepogu, Manoj and Akula, Sai Charitha and Yang, Jihan and Yang, Shusheng and Iyer, Adithya and Pan, Xichen and Wang, Austin and Fergus, Rob and LeCun, Yann and Xie, Saining , booktitle =. 2024 , doi = on Pith
  55. [75] Zhao, Yanpeng and Ding, Wentao and Li, Hongtao and Jia, Baoxiong and Zheng, Zilong , journal =. 2026 , url = on Pith
  56. [76] Vision in Action: Learning Active Perception from Human Demonstrations , author =. Proceedings of the 9th Conference on Robot Learning , volume =. 2025 , url = on Pith
Cited by 5 Pith papers

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment