Robotics · arXiv:2610.27613 v1 · submitted 20 Oct 2026 · 39 references
Embodied Task Agents Can Perform Fine Manipulation through Tactile In-Context Learning
The paper says
“Can embodied task agents use tactile experience to perform fine manipulation without task-specific model training? We investigate this capability through Tactile-ETA, a system that incorporates tactile experience into an embodied task agent through in-context learning.”
more of the abstract
Successful expert demonstrations show how contact changes during manipulation. We pair each recorded robot movement with its camera views and bilateral tactile observations. We keep tactile observations at the start and end of each movement and add two neighboring time points when the tactile images change markedly in between. The agent combines these demonstrations with current vision, tactile observations, and proprioception to select actions through general-purpose robot tools, without updating model parameters. Across four simulated manipulation tasks, adding tactile observations to the same two expert demonstrations increases average success from 15.75% to 24.00%, with current sensing and execution interfaces held fixed. Further experiments show that performance depends on the sensory content and representation of demonstrations, while comparing agent models on Lift Bottle reveals substantial differences under the same system configuration. On the real-robot tasks, demonstrations with aligned visual and tactile observations increase the task success rate from 0% to 60%. These findings demonstrate the feasibility of using tactile experience in context for fine manipulation and suggest a path toward extending embodied task agents to additional sensory inputs.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 21 Oct 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. Solid empirical study of tactile in-context learning for manipulation, but the main result is clouded by a missing control for context length. 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 39 references say.
The paper says, at §4.1, Table 1
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The main comparison (Visual-motion demos vs. Tactile-ETA) varies both the presence of tactile content in demonstrations and the total number of images provided in the context window. A control that adds an equal number of non-informative images (e.g., static mean tactile frames) to the Visual-motion demos is absent.
more of objection 1
Without this control, the observed gain of 8.25 percentage points cannot be unambiguously attributed to the tactile information rather than to the increased number of visual tokens per decision step. The content ablation (Table 2) and the Insert Tube result (where visual-motion demos add images over no demos but yield no gain) provide partial evidence, but they do not hold image count constant between the two compared conditions. This is a load-bearing issue for the central claim that tactile experience in demonstrations improves performance.
The paper says, at §4.3, Table 2 and Table 3
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
Table 3 reports that retaining all tactile frames (All frames) reduces success to 2%, compared to 36% for Selected and 27% for Endpoints. The paper states that performance depends on representation but does not analyze why all frames performs so poorly. A brief explanation (e.g., context-length constraints, noise, or distraction from irrelevant frames) would strengthen the ablation and clarify the design choices of the selection rule.
The paper says, at §D, Table 8
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
Increasing the number of demonstrations from two to four reduces average success from 24% to 19%, with sizable drops on Insert Tube and Lift Bottle. This result is reported without discussion of possible reasons (e.g., context window limits, increased ambiguity, or demonstration quality). A short analysis or hypothesis would improve the completeness of the study.
5 minor objections
The paper says, at §4.1, Table 1 footnote
Pith says: minor objection 1 of 5
The table caption says 'Bold denotes the best result in each row,' but the row for Lift Can shows 13 as best, which is bolded. However, Lift Can seems challenging for all conditions (max 13%), and the row average is also bolded. This is fine but might be clarified that 13% is indeed the maximum.
The paper says, at §3.2, Eq. (2)
Pith says: minor objection 2 of 5
The selection rule uses indices {0, n*-1, n*, N-1}. For clarity, specify that when n* = 1, the pair (0, n*-1) would duplicate the endpoint; deduplication is mentioned but could be stated explicitly here.
show 3 more
The paper says, at §4.5, Figure 6
Pith says: minor objection 3 of 5
The real-world results are based on only 5 trials per condition. While the effect is large, the small sample size should be acknowledged more prominently in the main text, not only in the appendix.
The paper says, at Appendix B, tactile selection parameters
Pith says: minor objection 4 of 5
The motion-change threshold and image-difference threshold are a priori parameters. The paper states they are fixed (Table 6), but a sensitivity analysis (e.g., varying the threshold by ±20%) would strengthen the claim that the selection rule is robust. This is a suggestion, not a requirement.
The paper says, at General
Pith says: minor objection 5 of 5
The paper would benefit from a brief discussion of computational cost: how many tokens the context uses and whether the agent's inference time increases with tactile observations.
What would settle it
Add a condition 'Visual–motion + dummy tactile' to the main simulation comparison (Table 1). In this condition, for each demonstration segment, include dummy tactile images (e.g., the mean tactile frame over the entire episode, computed per fingertip) in the same positions and counts as the real tactile observations in Tactile-ETA. Run 100 trials per task (Insert Tube, Lift Can, Lift Bottle, Pull Out Key). If the average success rate remains statistically indistinguishable from the Visual‑motion baseline (≈15.75%), the improvement is driven by tactile content; if it rises towards the Tactile-ETA result (≈24.00%), the improvement is largely an artifact of increased image count.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 21 Oct 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
This paper presents a clear system for adding tactile observations from demonstrations into a frozen embodied agent via in-context learning. The experimental design is careful: 100 trials per condition, four tasks, ablations over content and representation, and real-robot validation. The action-aligned segmentation and compact tactile selection (§3.2) are sensible engineering choices. The result—8.25 percentage point average gain from adding tactile to visual-motion demos—is modest but consistent.
The soft spot is the one flagged by the stress test. The central comparison (Table 1) confounds tactile content with the number of images in the context window. Visual–motion demos contain only demo vision frames; Tactile-ETA adds demo tactile frames. A control that adds an equal number of non-informative images to the visual–motion condition is absent. Without that control, part of the gain could come from giving the model more visual tokens per step. The authors have partial counterevidence: on Insert Tube, visual–motion demos produce no gain over the no-demo baseline, so simply adding images is not always enough. But that is indirect. The content ablations (Table 2) also do not keep image count constant because tactile-motion replaces vision images rather than adding to them. This does not invalidate the claim, but it weakens causal attribution.
Other concerns are minor: only two fixed demonstrations per task is a thin evaluation of robustness, and the improvement is small enough that chance or task selection could matter. The paper is honest about its limitations and provides extensive detail.
Who this is for: researchers working on in-context learning for robotics, especially those interested in multimodality. The paper is worth a serious referee because the experimental methodology is above average for the area. I would recommend conditional acceptance with a request for the missing control (dummy tactile frames) or at minimum a candid discussion of the confound.
The strongest objection
The central claim that adding tactile observations to demonstrations improves success (Table 1) rests on comparing Visual-motion demos (demo vision only) with Tactile-ETA (demo vision + demo tactile). This comparison varies both tactile content and the total number of images in the context. A direct control that adds an equal number of non-informative images (e.g., static mean tactile frames) to Visual-motion demos is absent. The content ablation (Table 2) and the Insert Tube task (where visual-motion demos add images but yield no gain) provide partial evidence that tactile content matters, but they do not hold image count constant: Tactile‑motion replaces vision images with tactile images rather than adding to them, so the number and type of images still differ. Without the missing control, the observed gain of 8.25 percentage points could be partially or entirely due to increased context length rather than to the specific contact information in the tactile images.
The reader's weakest_assumption correctly identifies the confound between tactile content and the number of images. The paper's internal controls (Table 2, Insert Tube result) partially address this but do not completely rule out a context‑length explanation. The proposed concrete test would settle the issue by holding image count constant while varying only the informativeness of the added images. Since the paper already acknowledges the concern (the reader gave CONDITIONAL, and the authors could respond by adding this control), the verdict should remain CONDITIONAL (that is, unchanged from the reader's recommendation). The strength of the paper's claims would be substantially improved if this control is included; without it, causal attribution remains uncertain. No further concerns are more load‑bearing: the results are internally consistent, the behavioral analyses support the qualitative interpretation, and the real‑robot results, though small, are consistent with the simulation findings.
The referee's report
If the central claim holds, the paper demonstrates a feasible method for extending embodied task agents to tactile information without retraining, which is a practical contribution for contact-rich manipulation. The system design—action-aligned multimodal demonstrations with compact tactile selection—is well-motivated and clearly described. The controlled simulation protocol (100 trials per condition, fixed demonstrations, same initial conditions) strengthens the evaluation. However, the reported gains are modest (8.25 percentage points average) and task-dependent; the real-robot results are limited to 3/5 successes. The significance is therefore moderate: the work provides a useful proof-of-concept and a reproducible evaluation framework, but the effect size and generality leave room for further improvement.
Pith's reading of the paper (a machine's summary, not the authors')
The paper proposes Tactile-ETA, a system that integrates tactile experience into an embodied task agent through in-context learning without updating model parameters. By constructing action-aligned demonstrations that pair robot motions with camera views and bilateral tactile observations, and by selecting compact tactile keyframes that capture contact changes, the agent uses these demonstrations alongside current sensor readings to choose actions. Across four simulated manipulation tasks, adding tactile observations to two visual-motion demonstrations increases average success from 15.75% to 24.00% with fixed sensing and execution interfaces. On real-robot wiping and pick-up tasks, a single multimodal demonstration raises success from 0% to 60%. The results demonstrate the feasibility of using tactile experience in context for fine manipulation, with performance depending on demonstration content, tactile representation, and the underlying agent model.
What carries the argument. Tactile-ETA and its tactile keyframe selection algorithm. The algorithm retains tactile observations at motion boundaries and, when a change threshold is met, adds one adjacent pair spanning the peak change per segment, limiting each fingertip to at most four observations. This compact representation captures intermediate contact transitions (e.g., a widening blue band during tube insertion) that endpoint-only sampling misses, while avoiding the context-length burden of full sequences. The agent conditions frozen model parameters on these demonstrations, current vision, tactile, and proprioception, selecting actions through general-purpose robot tools.
Core claim. Embodied task agents can use tactile experience provided through in-context learning to improve fine manipulation without task-specific training. The key evidence is that, across four simulated tasks, adding tactile observations from two expert demonstrations to the same visual-motion demonstrations raises average success from 15.75% to 24.00%, and on real-robot wiping and chip-pickup tasks a single tactile-augmented demonstration lifts success from 0% to 60%. This shows that contact information in demonstrations can guide action adjustments such as retreating and lateral correction after contact, which vision alone does not provide.
The premise it rests on. The reported improvement from adding tactile observations to demonstrations is causally attributed to the tactile content, but the comparison between Visual-motion demos (no demo tactile) and Tactile-ETA (demo tactile) does not control for the total number of images in the demonstration context—an additional non-informative image baseline is missing—so part of the gain could stem from the model receiving more visual tokens per decision rather than from tactile information.
If the paper is right (Pith's inferences, not the authors')
Demonstrated tactile experience can guide contact-sensitive adjustments during fine manipulation, such as retreating and lateral correction after detecting pronounced distortion.
The contribution of visual and tactile demonstration content is task-dependent: on Pull Out Key, tactile-motion demonstrations match full performance, while on Lift Bottle the complete representation works best.
How tactile frames are selected matters: retaining all frames reduces success to 2% on Lift Bottle, while the selective method achieves 36%, compared to 27% for endpoints-only.
Different agent models using the same demonstrations show large performance differences (0%–36% on Lift Bottle), indicating that context design must consider model capability.
The comparison between Visual-motion demos (no demo tactile) and Tactile-ETA (demo tactile) confounds tactile content with total image count, since no control adds an equal number of non-informative images to the visual-motion baseline. Part of the reported 8.25-point gain could be due to more context images rather than tactile information per se.
The tactile keyframe selection mechanism generalizes to other continuous observation modalities (e.g., force-torque, thermal) that exhibit within-movement transitions, providing a template for compact multi-modal in-context demonstrations.
Real-robot results, while positive, are limited to two tasks and five trials per condition; larger-scale evaluations with more tasks, more trials, and varied tactile sensors would test robustness of the approach.
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
Can embodied task agents use tactile experience to perform fine manipulation without task-specific model training? We investigate this capability through Tactile-ETA, a system that incorporates tactile experience into an embodied task agent through in-context learning. Successful expert demonstrations show how contact changes during manipulation. We pair each recorded robot movement with its camera views and bilateral tactile observations. We keep tactile observations at the start and end of each movement and add two neighboring time points when the tactile images change markedly in between. The agent combines these demonstrations with current vision, tactile observations, and proprioception to select actions through general-purpose robot tools, without updating model parameters. Across four simulated manipulation tasks, adding tactile observations to the same two expert demonstrations increases average success from 15.75% to 24.00%, with current sensing and execution interfaces held fixed. Further experiments show that performance depends on the sensory content and representation of demonstrations, while comparing agent models on Lift Bottle reveals substantial differences under the same system configuration. On the real-robot tasks, demonstrations with aligned visual and tactile observations increase the task success rate from 0% to 60%. These findings demonstrate the feasibility of using tactile experience in context for fine manipulation and suggest a path toward extending embodied task agents to additional sensory inputs.
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 · 2026-10-21
The authors' main claim, as the classifier picked it out 1.00
“These findings demonstrate the feasibility of using tactile experience in context for fine manipulation and suggest a path toward extending embodied task agents to additional sensory inputs.”
How the main claim is stated: asserted 0.72
What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.45 · human participants 0.05 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.83 · mentions code or data availability 0.03
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 |
|---|---|---|---|---|
| Can embodied task agents use tactile experience to perform fine manipulation without task-specific model training? | background 0.52 | 0.04 | 0.03 | unclear 0.75 |
| We investigate this capability through Tactile-ETA, a system that incorporates tactile experience into an embodied task agent through in-context learning. | contribution 0.80 | 0.90 | 0.06 | asserted 0.89 |
| Successful expert demonstrations show how contact changes during manipulation. | background 0.52 | 0.46 | 0.04 | asserted 1.00 |
| We pair each recorded robot movement with its camera views and bilateral tactile observations. | method 1.00 | 0.95 | 0.02 | asserted 0.99 |
| We keep tactile observations at the start and end of each movement and add two neighboring time points when the tactile images change markedly in between. | method 1.00 | 0.95 | 0.02 | asserted 0.97 |
| The agent combines these demonstrations with current vision, tactile observations, and proprioception to select actions through general-purpose robot tools, without updating model parameters. | method 0.99 | 0.88 | 0.10 | asserted 1.00 |
| Across four simulated manipulation tasks, adding tactile observations to the same two expert demonstrations increases average success from 15.75% to 24.00%, with current sensing and execution interfaces held fixed. | result 1.00 | 0.97 | 0.15 | asserted 1.00 |
| Further experiments show that performance depends on the sensory content and representation of demonstrations, while comparing agent models on Lift Bottle reveals substantial differences under the same system configuration. | result 0.98 | 0.95 | 0.06 | asserted 1.00 |
| On the real-robot tasks, demonstrations with aligned visual and tactile observations increase the task success rate from 0% to 60%. | result 1.00 | 0.96 | 0.14 | asserted 1.00 |
| These findings demonstrate the feasibility of using tactile experience in context for fine manipulation and suggest a path toward extending embodied task agents to additional sensory inputs. | interpretation 0.99 | 0.87 | 0.91 | asserted 0.72 |
What kind of paper the classifier reports this is: method or tool 0.67 · 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, 2026-10-21: 1 request, 7,444 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 (7)
Showing 6 of the paper's 7 figures. Figures are the authors', reproduced from the paper.
References (39)
- [1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, ... on PithPith's reviewarXiv
- [2] Brenna D. Argall, Sonia Chernova, Manuela Veloso, and Brett Browning. A survey of robot learning from demonstration.Robotics and Autonomous Systems, 57(5):469–483, 2009. ISSN 0921-8890. doi: https://doi.org/10.1016/ j.robot.2008.10.024. URLhttps://www.sciencedirect.com/science... on Pith
- [3] Jianxin Bi, Kevin Yuchen Ma, Ce Hao, Mike Zheng Shou, and Harold Soh. Vla-touch: Enhancing vision-language- action models with dual-level tactile feedback, 2025. URLhttps://arxiv.org/abs/2507.17294 on PithPith's reviewarXiv
- [4] Baijun Chen, Weijie Wan, Tianxing Chen, Xianda Guo, Congsheng Xu, Yuanyang Qi, Haojie Zhang, Longyan Wu, Tianling Xu, Zixuan Li, Yizhe Wu, Rui Li, Xiaokang Yang, Ping Luo, Wei Sui, and Yao Mu. Univtac: A unified simulation platform for visuo-tactile manipulation data generation,... on PitharXiv
- [5] Kaiyuan Chen, Shuangyu Xie, Letian Fu, Justin Yu, William Pacini, Sandeep Bajamahal, Hudson Kim, Jaimyn Drake, Daehwa Kim, Haoru Xue, Jonathan Francis, Christian Juette, Peter Schaldenbrand, Muhammet Yunus Seker, Ruwan Wickramarachchi, Uksang Yoo, Guanzhi Wang, Adithyavairavan... on PithPith's reviewarXiv
- [6] Kejia Chen, Zheng Shen, Yue Zhang, Lingyun Chen, Fan Wu, Zhenshan Bing, Sami Haddadin, and Alois Knoll. Lemmo-plan: Llm-enhanced learning from multi-modal demonstration for planning sequential contact-rich manipulation tasks, 2025. URLhttps://arxiv.org/abs/2409.11863 on PithPith's reviewarXiv
- [7] YanzheChen,ZechenBai,ZhijunCao,WenzhengZeng,KevinQinghongLin,YiqiLin,GuoqiangLiang,KevinYuchen Ma, Qiming Huang, and Mike Zheng Shou. Show-harness: Just a vlm agent can play robots, 2026. URL https://arxiv.org/abs/2609.10522 on PithPith's reviewarXiv
- [8] Yitong Chen, Zezheng Huai, Sixian Li, Yubang Wang, Haozhe Zhang, Yifei Zhang, Hechang Chen, Jingjing Gong, Yu-Gang Jiang, and Xipeng Qiu. Eta: A new agentic paradigm for embodied tasks, 2026. URLhttps: //arxiv.org/abs/2608.03924 on PithPith's reviewarXiv
- [9] Dongzhou Cheng, Taoran Yi, Ye Fang, Xingwu Zhang, Fan Feng, Yixuan Li, Gengxiong Zhuang, Rongze Wang, Shuai Yang, Wei Song, Weizhi Xue, Minyan Wu, Jie Gui, Jiaqi Wang, and Tong Wu. In-context robot learning with vlm agents, 2026. URLhttps://arxiv.org/abs/2609.19138 on PithPith's reviewarXiv
- [10] 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 next-token prediction, 2024. URLhttps://arxiv.org/ abs/2408.15980 on PithPith's reviewarXiv
- [11] Letian Fu, Justin Yu, Karim El-Refai, Ethan Kou, Haoru Xue, Huang Huang, Wenli Xiao, Guanzhi Wang, Dantong Niu, Fei-Fei Li, Guanya Shi, Jiajun Wu, Shankar Sastry, Yuke Zhu, Ken Goldberg, and Linxi "Jim" Fan. Cap- x: A framework for benchmarking and improving coding agents for ... on PithPith's reviewarXiv
- [12] Meng-Hao Guo, Zhe-Han Mo, Jia-Jun Wang, Yi Zhang, Kejin Wang, Yi-Xuan Deng, Jia-Peng Zhang, Yongming Rao, and Shi-Min Hu. Transferring the intelligence of vlms to robotic control, 2026. URLhttps://arxiv.org/abs/ 2609.22966. 11 on PithPith's reviewarXiv
- [13] Irmak Guzey, Yinlong Dai, Ben Evans, Soumith Chintala, and Lerrel Pinto. See to touch: Learning tactile dexterity through visual incentives, 2023. URLhttps://arxiv.org/abs/2309.12300 on PithPith's reviewarXiv
- [14] Peng Hao, Chaofan Zhang, Dingzhe Li, Xiaoge Cao, Xiaoshuai Hao, Shaowei Cui, and Shuo Wang. Tla: Tactile- language-action model for contact-rich manipulation, 2025. URLhttps://arxiv.org/abs/2503.08548 on PithPith's reviewarXiv
- [15] Nick Heppert, Max Argus, Tim Welschehold, Thomas Brox, and Abhinav Valada. Ditto: Demonstration imitation by trajectory transformation. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), page 7565–7572. IEEE, October 2024. doi: 10.1109/iros58592... on Pithdoi:10.1109/iros58592.2024.10801982
- [16] CarolinaHiguera,AkashSharma,ChaithanyaKrishnaBodduluri,TaoshaFan,PatrickLancaster,MrinalKalakrishnan, Michael Kaess, Byron Boots, Mike Lambeta, Tingfan Wu, and Mustafa Mukadam. Sparsh: Self-supervised touch representations for vision-based tactile sensing. In Pulkit Agrawal, O... on Pith
- [17] Haoxu Huang, Fanqi Lin, Yingdong Hu, Shengjie Wang, and Yang Gao. Copa: General robotic manipulation through spatial constraints of parts with foundation models, 2024. URLhttps://arxiv.org/abs/2403.08248 on PithPith's reviewarXiv
- [18] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner monologue: Embodied reasoning th... on PithPith's reviewarXiv
- [19] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models, 2023. URLhttps://arxiv.org/abs/2307.05973 on PithPith's reviewarXiv
- [20] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation, 2024. URLhttps://arxiv.org/abs/2409.01652 on PithPith's reviewarXiv
show 19 more
- [21] Yuzhe Huang, Pei Lin, Wanlin Li, Daohan Li, Jiajun Li, Jiaming Jiang, Chenxi Xiao, and Ziyuan Jiao. Taf- vla: Tactile-force alignment in vision-language-action models for force-aware manipulation, 2026. URLhttps: //arxiv.org/abs/2601.20321 on PitharXiv
- [22] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. Vima: General robot manipulation with multimodal prompts, 2023. URL https://arxiv.org/abs/2210.03094 on PithPith's reviewarXiv
- [23] 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, 2023. URLhttps://arxiv.org/abs/2209.07753 on PithPith's reviewarXiv
- [24] Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-world robotic manipulation through mark-based visual prompting, 2024. URLhttps://arxiv.org/abs/2403.03174 on PithPith's reviewarXiv
- [25] Runyu Lu, Yubo Wu, Ethan Kou, Letian Fu, Wenli Xiao, Ajay Mandlekar, Yinzhen Xu, Guanya Shi, Ken Goldberg, Ang Chen, Mosharaf Chowdhury, Yuke Zhu, Linxi "Jim" Fan, and Guanzhi Wang. Aspire: Agentic /skills discovery for robotics, 2026. URLhttps://arxiv.org/abs/2607.00272 on PithPith's reviewarXiv
- [26] Achu Menon, Sravanthi Machcha, Sabrina Zou, Tzu Kit Chan, and Jay Chooi. Gpt-6 astra on robotic manipulation. https://openai.robocurve.org/gpt-6-astra/, September 2026. Robocurve, accessed September 26, 2026 on Pith
- [27] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, Quan Vuong, Tingnan Zhang, Tsang-Wei Edward Lee, Kuang-Huei Lee, Peng Xu, Sean Kirmani, YukeZhu,AndyZeng,KarolHausman,NicolasHeess,ChelseaFinn,SergeyL... on PithPith's reviewarXiv
- [28] NormanDiPaloandEdwardJohns. Dinobot: Robotmanipulationviaretrievalandalignmentwithvisionfoundation models, 2024. URLhttps://arxiv.org/abs/2402.13181 on PithPith's reviewarXiv
- [29] Norman Di Palo and Edward Johns. Keypoint action tokens enable in-context imitation learning in robotics, 2024. URLhttps://arxiv.org/abs/2403.19578. 12 on PithPith's reviewarXiv
- [30] Jiayi Su, Yixin Zheng, Mi Yan, Li Yi, Zhizheng Zhang, and He Wang. GPT 6 Astra as an em- bodied policy. Technical report and code, 2026. URL https://github.com/anonymous-report-421/ eval-of-gpt-6-astra-as-policy on Pith
- [31] Vitalis Vosylius and Edward Johns. Instant policy: In-context imitation learning via graph diffusion, 2025. URL https://arxiv.org/abs/2411.12633 on PithPith's reviewarXiv
- [32] Zidan Wang, Rui Shen, and Bradly Stadie. Wonderful team: Zero-shot physical task planning with visual llms, 2025. URLhttps://arxiv.org/abs/2407.19094 on PithPith's reviewarXiv
- [33] Wenli Xiao, Jia Xie, Tonghe Zhang, Haotian Lin, Letian "Max" Fu, Haoru Xue, Jalen Lu, Yi Yang, Cunxi Dai, Zi Wang, Jimmy Wu, Guanzhi Wang, S. Shankar Sastry, Ken Goldberg, Linxi "Jim" Fan, Yuke Zhu, and Guanya Shi. Enpire: Agentic robot policy self-improvement in the real worl... on PithPith's reviewarXiv
- [34] Congsheng Xu, Qiaochu Yang, Fangyuan Shi, Yifan Han, Baijun Chen, Yiming Wang, Haonan Zhao, Daolin Ma, Xiaokang Yang, and Hesheng Wang. Vt-muse: Multimodal unified sequential visuotactile representation learning for manipulation, 2026. URLhttps://arxiv.org/abs/2608.21290 on PithPith's reviewarXiv
- [35] Yida Yin, Zekai Wang, Yuvan Sharma, Dantong Niu, Trevor Darrell, and Roei Herzig. In-context learning enables robot action prediction in llms, 2025. URLhttps://arxiv.org/abs/2410.12782 on PithPith's reviewarXiv
- [36] Samson Yu, Kelvin Lin, Anxing Xiao, Jiafei Duan, and Harold Soh. Octopi: Object property reasoning with large tactile-language models, 2024. URLhttps://arxiv.org/abs/2405.02794 on PithPith's reviewarXiv
- [37] Samson Yu, Kelvin Lin, and Harold Soh. Demonstrating the octopi-1.5 visual-tactile-language model, 2025. URL https://arxiv.org/abs/2507.09985 on PithPith's reviewarXiv
- [38] Chengbo Yuan, Zicheng Zhang, Mingjie Zhou, Wendi Chen, Yi Wang, Zhuoyang Liu, Dantong Niu, Shuo Wang, Hui Zhang, Wenkang Zhang, Yingdong Hu, Yuanqing Gong, Wanli Xing, Chuan Wen, Cewu Lu, Kaifeng Zhang, and Yang Gao. Ftp-1: A generalist foundation tactile policy across tactile... on PithPith's reviewarXiv
- [730] The figure displays the endpoints and step 730, with synchronized wrist-camera and right-fingertip observations. Green annotations mark the same tactile image region in each column. The wider blue band in the intermediate frame illustrates information omitted by endpoint-only ... on Pith
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