Pith. sign in

REVIEW 5 major objections 6 minor 16 cited by

Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that fusing tactile sensing into a VLA model activates its latent physical knowledge and enables zero-shot force generalization.

desk verdict A useful tactile-VLA systems paper with a genuinely new zero-shot force-language evaluation, but the physical-semantics claim is undercut by uncalibrated force labels and missing ablations. read the letter →

arxiv 2507.09160 v1 pith:MMR3NLBD submitted 2025-07-12 cs.RO cs.LG

Jialei Huang , Shuo Wang , Fanqi Lin , Yihang Hu , Chuan Wen , Yang Gao This is my paper
classification cs.ROcs.LG
keywords tactilesensingvision-language-actionmodelforcecontrolcontact-richmanipulationzero-shotgeneralizationimitationlearningchain-of-thoughtreasoninghybridposition-forcecontroller
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Tactile-VLA is a framework for adding tactile sensing to vision-language-action models so that a robot can control not only what to do but how hard to do it. The paper's central claim is that the VLM's prior knowledge already contains semantic understanding of physical interaction, and that connecting this knowledge to tactile sensors with only a few demonstrations activates it. If correct, a robot can generalize force-related language like 'softly' or 'hard' to tasks never trained with force labels, choose gentle or firm grasps for objects it has never seen, and infer from contact feedback that it needs to press harder. The payoff is a path to contact-rich manipulation that does not require large force-labeled datasets.

What carries the argument

The load-bearing object is the tactile-aware action expert: an output head that predicts an augmented action vector $(P_{target}, F_{target})$ from a token-level fusion of vision, language, and tactile tokens in a pretrained VLA transformer, trained with a Conditional Flow Matching objective. A hybrid position-force controller then translates $F_{target}$ into positional adjustments according to force error $\Delta F = F_{target} - F_{measured}$, active only when $\lVert \Delta F \rVert$ exceeds a threshold; this separates the net external force, regulated by gripper Cartesian position, from the internal grasp force, regulated by gripper width. The CoT variant adds explicit language reasoning over tactile feedback at fixed intervals, so the model can diagnose a failure and re-plan the next force command.

What would settle it

Compare the $F_{target}$ values stored during teleoperation with an independent, calibrated force measurement at the gripper on the same objects; systematic disagreement would show the model learned a labeling artifact rather than physical interaction.

Watch

Extended reading notes

Core claim

Tactile-VLA's central claim is that a pretrained vision-language model already encodes semantic knowledge of physical interaction, and this knowledge can be activated for control by fusing tactile signals into the model and finetuning on a small demonstration dataset. The action expert is trained to emit both a target position and a target contact force; a hybrid position-force controller executes those targets by turning force error into small position corrections, with external contact force and internal grasp force regulated by independent channels. The paper reports three zero-shot transfer results supporting the claim: force adverbs learned during USB insertion modulate applied force during charger insertion, unseen objects receive grasp forces that match their apparent fragility, and chain-of-thought reasoning over tactile feedback lets the model increase force after a failed wiping attempt on a blackboard. The intended conclusion is that VLAs can be grounded in physical interaction without force-labeled data at scale.

Load-bearing premise

The recorded force targets in the demonstrations must faithfully represent the physical interaction the operator intended, because the model learns its force-language mapping from those values and no calibration or consistency check is reported.

Editorial extensions

If this is right

  • Force adverbs trained on one insertion task transfer to a second insertion task with zero force-labeled demonstrations, changing the applied force from soft to hard.
  • An unseen object is grasped with a force appropriate to its visible fragility, so a pitaya is lifted gently and an iron cube firmly.
  • A robot that initially fails to wipe a novel blackboard can reason over its tactile feedback, raise its downward force, and complete the wipe.
  • These abilities are obtained with roughly one hundred demonstrations per behavior, pointing to small-data finetuning of generalist VLAs.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the claimed force semantics genuinely live in the pretrained VLM, the same tactile-activation recipe should work for other physical quantities such as compliance, friction, or torque limits; the paper does not test those, but the mechanism is unchanged.
  • Because the controller separates net external force from internal grasp force, language could in principle modulate the two channels independently, for example 'hold firmly but push gently'; that would be a direct next experiment.
  • A decisive ablation would detach the tactile encoder while keeping the language conditioning; if force differentiation survives, the VLM alone carries the semantics, and if it collapses, the tactile stream is the real carrier rather than an activator of prior knowledge.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper presents Tactile-VLA, a vision-language-action policy that fuses tactile readings into a pretrained VLM (Gemma/pi0) and outputs both target position and target force; a hybrid position-force controller executes these targets, and a CoT variant performs explicit reasoning over tactile feedback. Experiments on USB/charger insertion, tabletop grasping, and board wiping claim zero-shot generalization of force adverbs, object-appropriate grasping forces for unseen objects, and adaptive force correction on a novel blackboard. The central claim is that a few demonstrations can activate latent physical semantics in a VLM to achieve zero-shot force generalization in contact-rich tasks.

Significance. The framework is plausible and addresses an important gap: grounding VLA outputs in physical force rather than only in motion. If the central claim were fully supported, the result would be valuable. The paper also has strengths: it provides an explicit action-space formulation with force targets, a clear controller in Eq. (2), and real-robot evaluations. However, the current evidence is insufficient to establish the headline mechanism: force labels are unvalidated, trial counts and confidence intervals are missing, no ablation isolates the tactile modality or the pretrained backbone, and concurrent tactile-VLA methods are not compared. The work is better viewed as a promising system paper than as a demonstrated discovery about VLM priors.

major comments (5)
  1. [Section 2.4 / Eq. (2)] The force measurement chain is not validated. Section 2.4 says the UMI gripper was augmented with tactile sensors so that operators can 'directly sense contact dynamics', but the paper reports no calibration of these sensors, no conversion from sensor readings to Newtons, no verification that recorded Ftarget values match actual contact forces, and no statistics across operators or trials. Because Eq. (2) makes the controller actively track Ftarget, the 'applied force' values reported in Table 2 and Figure 6 are largely the model's own output, not an independent measurement of physical interaction. The zero-shot semantic-generalization claim in Section 3.2 therefore rests on an unverified mapping from sensor readings to force. Please calibrate the sensors against a reference force/torque sensor, report measured force traces during evaluation, and provide cross-operator and cross-trial consistency statistics.
  2. [Tables 1, 2, 4 and Figure 6] The central quantitative claims are presented without trial counts or confidence intervals. Table 1 reports success rates without stating n; Table 2 reports two-decimal force values without error bars; Table 4 has success rates with no n or CIs; Figure 6 states only 5 trials per condition. Given that the headline results include close comparisons such as 80% versus 75% in Table 4 and force orderings such as 4.68 versus 9.13 N in Table 2, the reader cannot assess whether the differences are reliable. Please report the number of trials for every condition, per-trial values or distributions, and appropriate confidence intervals or significance tests.
  3. [Section 3.1 (Baselines)] No ablation isolates the claimed source of generalization. The baselines pi0 and pi0-fast differ from Tactile-VLA in several ways at once: they lack the tactile encoder, lack the force-augmented action space, and are not coupled to the hybrid controller of Eq. (2). The observed gains and the force differentiation in Table 2 could therefore be due to the controller or the action representation rather than to the VLM's prior knowledge invoked in Section 5. Please add ablations that remove the tactile tokens, replace the pretrained VLM backbone with a randomly initialized one, and remove the force action head, so that the contribution of the pretrained physical semantics is actually identified.
  4. [Section 4 and Section 3.1] The closest prior tactile-VLA methods are not compared. Related work names FuSe and ForceVLA as concurrent works with different design choices, but all experiments compare only against pi0 and pi0-fast. Since the paper claims to outperform standard VLA baselines in tactile settings, and the central novelty is the particular fusion design, at least one head-to-head comparison, or a clear statement of why such a comparison is not possible, is needed before claiming superiority over alternative tactile-VLA designs.
  5. [Section 2.3 and Section 3.4] The CoT experiment does not establish that reasoning is driven by tactile feedback. Training data includes corrective annotations such as 'The force was too light... Now trying with 5N', and Section 2.3 states that CoT is triggered at fixed intervals. The blackboard success could therefore arise from copying a learned corrective response to a new visual context, or from scheduled re-planning, rather than from interpreting tactile signals. Please add control conditions without tactile input, without force feedback, and with CoT triggered only by actual failure detection; also report success rates with trial counts and confidence intervals.
minor comments (6)
  1. [Eq. (1)] Please define Tt and the history length H explicitly; the notation [Tt-H+1, ..., Tt] is ambiguous about how many tactile measurements are included in the fused token.
  2. [Figure 2] The architecture diagram has unclear and overlapping labels, including 'ViT Tokenizer Encoder Encoder', and the dashed block for the CoT variant is difficult to read.
  3. [Figure 5(b)] The label 'Piyata' appears to be a typo, likely 'Pitaya' or 'Pinata', and should be corrected for consistency with the text.
  4. [Table 1] The caption should state the number of evaluation trials used for each success rate and clarify that the Charger column is the zero-shot condition.
  5. [Section 2.4 and abstract] The abstract and introduction emphasize 'only a few demonstrations', but Section 3.1 reports 100 demonstrations per task for the main tasks; please clarify what 'few' means relative to the baseline training data.
  6. [References] Several references are incomplete or inconsistently formatted, including entries for Hu et al., Huang et al., and Yu et al.; please ensure all entries have full author lists and publication details.

Circularity Check

0 steps flagged · score 0.0 of 10

No equation-level circularity: the zero-shot force outputs are evaluated on held-out adverbs, objects, and tasks rather than fitted constants; the force-label calibration gap is a measurement-validity concern, not a circular derivation.

full rationale

I checked every load-bearing step in the claimed derivation chain. The model is trained end-to-end to output Ftarget from demonstrations, and the hybrid controller in Eq. (2) tracks that target so the measured applied force will largely equal the model's own setpoint; however, the paper's generalization claims are not reduced to the training labels by construction. The key evaluations are on held-out inputs: force adverbs not seen during training (e.g., 'gently', 'firmly', 'harder'), a zero-shot charger task whose demonstrations contained no force-language supervision, out-of-domain objects never grasped during training, and a blackboard wiping scenario never experienced during training. These are genuine generalization tests, not re-statements of fitted parameters. The paper's self-citations, such as OneTwoVLA in the CoT-related work discussion, are peripheral and do not carry the central premise. No uniqueness theorem, ansatz, or definitional equivalence is imported from the authors' prior work to force the conclusion. The absence of tactile-sensor calibration, the lack of a sensor-count-to-Newton conversion procedure, and the absence of cross-operator consistency checks are important validity risks for the physical-semantics interpretation, but they do not make the derivation circular: they affect whether the force labels are trustworthy, not whether the reported outputs are logically identical to the inputs. The central claim that a VLM's prior knowledge contains physical interaction semantics is an interpretation of observed zero-shot transfer rather than a construct that is true by definition. Overall, no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central contribution rests on two categories of unproven inputs: the claim that a pretrained VLM already encodes physical interaction semantics, and the assumption that the tactile data collection and force control pipeline produce trustworthy force targets. No new theoretical entities are introduced.

free parameters (3)
  • Controller gain matrix K = not reported
    Equation 2 converts force error into position correction; its magnitude determines how aggressively force targets are tracked. No values or tuning procedure are given.
  • Force-error threshold tau = not reported
    Equation 2 only applies position corrections when the force error exceeds tau. The threshold affects contact smoothness and force accuracy; no value is reported.
  • CoT triggering interval = not reported
    Section 2.3 says reasoning is triggered at fixed intervals, but the interval length is not specified. It controls how often the model can adapt its force strategy.
assumptions (4)
  • ad hoc to paper Pretrained VLM (Gemma 2.6B) retains reusable semantic knowledge about physical interaction after fine-tuning for action generation.
    This is the central hypothesis. The paper does not ablate the pretrained backbone or compare with a randomly initialized model, so the presence of this prior is assumed rather than demonstrated.
  • domain assumption Expert demonstrations collected with the augmented UMI gripper provide accurate, synchronized force targets and tactile readings.
    Section 2.4 describes the setup but reports no calibration, synchronization error bounds, or inter-operator consistency check.
  • domain assumption The hybrid position-force controller stably tracks target forces in the evaluated tasks with the chosen K and tau.
    Section 2.2 assumes impedance-style position adjustment is sufficient; no stability analysis or controller tuning details are provided.
  • domain assumption Success metrics, such as insertion success, grasp without deformation, and wipe cleanliness, are reliable and consistently applied.
    Success is judged visually or by deformation, with no rubric or inter-annotator agreement reported; OOD object categories are assigned by the authors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization." pith.science (2026). https://pith.science/paper/MMR3NLBD

@misc{pith2026250709160,
  author       = {Pith},
  title        = {Pith review of: Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MMR3NLBD}},
  note         = {Machine review of arXiv:2507.09160}
}
read the original abstract

Vision-Language-Action (VLA) models have shown remarkable achievements, driven by the rich implicit knowledge of their vision-language components. However, achieving generalist robotic agents demands precise grounding into physical interactions, especially in contact-rich scenarios where fine-grained force control is essential. We advance VLAs' implicit knowledge beyond identifying what to do, towards guiding how to physically interact with real world. This paper introduces Tactile-VLA, a novel framework that deeply fuses vision, language, action, and tactile sensing. This framework incorporates a hybrid position-force controller to translate the model's intentions into precise physical actions and a reasoning module that allows the robot to adapt its strategy based on tactile feedback. Experiments demonstrate Tactile-VLA's effectiveness and generalizability in three key aspects: (1) enabling tactile-aware instruction following, (2) utilizing tactile-relevant commonsense, and (3) facilitating adaptive tactile-involved reasoning. A key finding is that the VLM's prior knowledge already contains semantic understanding of physical interaction; by connecting it to the robot's tactile sensors with only a few demonstrations, we can activate this prior knowledge to achieve zero-shot generalization in contact-rich tasks.

Figures

Figures reproduced from arXiv: 2507.09160 by the authors.

Figure 1
Figure 1. Key capabilities of Tactile-VLA. (a) Generalizing language-based force control: It applies force modifiers (e.g., ‘gently’) learned from one task to a new task for which it only learned the motion. (b) Applying implicit common sense: The model automatically uses appropriate firm or gentle grasps for unseen objects without explicit force commands. (c) Reasoning to overcome fail￾ure: It generalizes reasoning learned f… view at source ↗
Figure 2
Figure 2. Overview of the Tactile-VLA architecture. Vision, language, tactile, and proprioceptive [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The working pro￾cess of Tactile-VLA-CoT on Wiping the Board task. However, unlike classic impedance control which aims for passive compliance, our objective is the active tracking of a target force. The controller measures the force error ∆F = Ftarget − Fmeasured, which is used to compute a corrective positional adjustment only when its magnitude ∥∆F∥ exceeds a predefined threshold τ to en￾hance operational smoothne… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Data collection setup Accurate and semantically aligned tactile data is critical for training agents in contact-rich scenarios. Conventional teleop￾eration is insufficient for this purpose, as the human operator typically lacks direct force feedback. A policy collected…
Figure 5
Figure 5. Figure 5: (a) The charger insertion and extraction task. (b) A selection of objects for the tabletop [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Applied grasping force for various objects, categorized by hardness and whether they are [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Generalizing Wiping Strategies through Tactile-Involved Reasoning. (a) The model is first [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 16 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Feeling the Unexpected: ResTacVLA for Contact-Rich Manipulation via Residual Tactile Representation

    cs.RO 2026-07 conditional novelty 6.5 of 10

    Residual tactile representations plus surprise-aware gating let VLA policies master contact-rich robot tasks that vision-only models fail.

  2. {\tau}: Learning Touch-Augmented Vision-Language-Action Models from Future Visual Supervision

    cs.RO 2026-07 conditional novelty 6.0 of 10

    Action-conditioned JEPA-style future-visual latent prediction yields dynamics-aware tactile tokens that lift contact-rich VLA success rates from ~30% to ~70% average on four real tasks.

  3. $N_0$-TWAM: Scaling Tactile-Native World-Action Model for Contact-Rich Manipulation

    cs.RO 2026-07 conditional novelty 6.0 of 10

    A scaled tactile-native world-action model jointly predicts vision, touch, and action and outperforms vision-only baselines on contact-rich sim and real robot tasks.

  4. $N_0$-VTLA: Scaling Vision-Tactile-Language-Action Model with Latent Tactile Tokens

    cs.RO 2026-07 conditional novelty 6.0 of 10

    A VLA with predictive latent tactile tokens pretrained on large-scale visuo-tactile data, plus ALTER offline advantage labeling, leads contact-rich real and sim benchmarks.

  5. FORGE-plus: Force-Budgeted Recovery for Contact-Rich Assembly with a Frozen LLM Supervisor

    cs.RO 2026-07 conditional novelty 6.0 of 10

    With a hidden per-episode breaking force, an LLM-set force ceiling plus force-signature recovery achieves 256/256 clean insertions on fragile and robust parts and resolves 40–64% of injected jams in simulation.

  6. FELT: Generating Tactile Signals from Vision for Visuo-Tactile Manipulation

    cs.RO 2026-07 conditional novelty 6.0 of 10

    FELT predicts finger pressure maps from RGB images and uses them or their learned features to improve manipulation policies without real tactile sensors at deployment.

  7. Representation-Aligned Tactile Grounding for Contact-Rich Robotic Manipulation

    cs.RO 2026-07 conditional novelty 6.0 of 10

    Future tactile prediction applied to intermediate action-expert features, rather than visual-language or final-action features, improves contact-rich manipulation in SmolVLA and π0.

  8. SoftVTBench: A Safety-Aware Visuo-Tactile Benchmark for Physically Constrained Robotic Manipulation of Deformable Objects

    cs.RO 2026-07 conditional novelty 6.0 of 10

    Success-only metrics overstate deformable-manipulation performance; tactile sensing raises Safety Success (e.g. 21.4%→35.6% on Object-Soft) while Goal Success stays comparable.

  9. TACO: TActile World Model as a Self-COrrector forScalable VLA Post-Training

    cs.RO 2026-07 conditional novelty 6.0 of 10

    A tactile-aware world model recognizes failure-adjacent contact states, imagines local visuo-tactile corrections, and post-trains VLAs with knowledge insulation, raising average success by 44% over the base policy.

  10. Tactile Modality Fusion for Vision-Language-Action Models

    cs.RO 2026-03 conditional novelty 6.0 of 10

    A FiLM-based tactile fusion method that conditions VLA visual features on frozen pretrained touch embeddings improves real-robot insertion success, speed, and force control relative to vision-only and concatenation baselines.

  11. Learning to Feel the Future: DreamTacVLA for Contact-Rich Manipulation

    cs.RO 2025-12 unverdicted novelty 6.0 of 10

    DreamTacVLA grounds VLA models in contact physics by aligning multi-scale vision-tactile inputs and predicting future tactile states, reaching up to 95% success on contact-rich tasks.

  12. LightTact: A Visual-Tactile Fingertip Sensor for Deformation-Independent Contact Sensing

    cs.RO 2025-12 conditional novelty 6.0 of 10

    A fingertip camera sensor uses a light-blocking wedge so that only true contact pixels brighten, enabling deformation-free contact detection with liquids, soft materials, and rigid objects.

  13. Prior Reinforce: Goal-Conditioned Dynamic Manipulation with Limited Trials

    cs.RO 2025-05 conditional novelty 6.0 of 10

    Prior Reinforce adapts a few demonstration motions to new goals in dynamic manipulation by learning a diffusion motion prior and refining a low-dimensional condition via Bayesian optimization, reaching new goals in un...

  14. TacWAM: Anchor-Guided World Action Model with Mechanics-Aware Tactile Prediction

    cs.RO 2026-07 conditional novelty 5.0 of 10

    Mechanics-aware future tactile prediction plus history and train-time isolation of future tokens raises real-robot contact-rich success from ~37.5% to 75% average.

  15. ViTacWorld: Scaling Visuo-Tactile World Models for Contact-Rich Robot Manipulation

    cs.RO 2026-07 conditional novelty 5.0 of 10

    An action-conditioned visuo-tactile world model generates synthetic camera-plus-touch rollouts that, mixed with real demonstrations, improve downstream contact-rich manipulation policies.

  16. HapticVLA: Contact-Rich Manipulation via Vision-Language-Action Model without Inference-Time Tactile Sensing

    cs.RO 2026-03 conditional novelty 5.0 of 10

    A vision-language-action robot policy that claims tactile-aware manipulation without tactile sensors at inference via reward-weighted flow matching and action distillation.

Reference graph

Works this paper leans on

24 extracted references · 7 canonical work pages · cited by 16 Pith papers

  1. [1]

    Rt-h: Action hierarchies using language

    Suneel Belkhale, Tianli Ding, Ted Xiao, Pierre Sermanet, Quon Vuong, Jonathan Tompson, Yevgen Chebotar, Debidatta Dwibedi, and Dorsa Sadigh. Rt-h: Action hierarchies using language. arXiv preprint arXiv:2403.01823,

    arXiv

  2. [3]

    Rt-1: Robotics transformer for real-world control at scale

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817,

    arXiv

  3. [5]

    Tactile-rl for insertion: Generalization to objects of unknown geometry

    Siyuan Dong, Devesh K Jha, Diego Romeres, Sangwoon Kim, Daniel Nikovski, and Alberto Ro- driguez. Tactile-rl for insertion: Generalization to objects of unknown geometry. In 2021 IEEE International Conference on Robotics and Automation (ICRA) , pp. 6437–6443. IEEE,

    work page 2021

  4. [9]

    Vima: General robot manipulation with multimodal prompts

    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. In NeurIPS 2022 F oundation Models for Decision Making Workshop. Joshua Jones, Oier Mees, Carmelo Sferrazza, Kyle Stachowicz, Pieter Abbeel, and Sergey Levine. ...

    arXiv 2022

  5. [11]

    Learning visuotactile skills with two multifingered hands

    Toru Lin, Yu Zhang, Qiyang Li, Haozhi Qi, Brent Yi, Sergey Levine, and Jitendra Malik. Learning visuotactile skills with two multifingered hands. arXiv preprint arXiv:2404.16823,

    arXiv

  6. [12]

    Forcemimic: Force-centric imitation learning with force-motion capture system for contact-rich manipulation.arXiv preprint arXiv:2410.07554,

    Wenhai Liu, Junbo Wang, Yiming Wang, Weiming Wang, and Cewu Lu. Forcemimic: Force-centric imitation learning with force-motion capture system for contact-rich manipulation.arXiv preprint arXiv:2410.07554,

    arXiv

  7. [13]

    Dense tactile force estimation using gelslim and inverse fem

    Daolin Ma, Elliott Donlon, Siyuan Dong, and Alberto Rodriguez. Dense tactile force estimation using gelslim and inverse fem. In 2019 International Conference on Robotics and Automation (ICRA), pp. 5418–5424. IEEE,

    work page 2019

  8. [14]

    Learning precise, contact-rich manipulation through uncalibrated tactile skins

    Venkatesh Pattabiraman, Yifeng Cao, Siddhant Haldar, Lerrel Pinto, and Raunaq Bhirangi. Learning precise, contact-rich manipulation through uncalibrated tactile skins. In CoRL 2024 Workshop CoRoboLearn: Advancing Learning for Human-Centered Collaborative Robots . Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees,...

    arXiv 2024

Show all 24 references
  1. [16]

    Yell at your robot: Improving on-the-fly from language corrections

    Lucy Xiaoyang Shi, Zheyuan Hu, Tony Z Zhao, Archit Sharma, Karl Pertsch, Jianlan Luo, Sergey Levine, and Chelsea Finn. Yell at your robot: Improving on-the-fly from language corrections. arXiv preprint arXiv:2403.12910,

    arXiv

  2. [17]

    Enhancing physical human-robot interaction: Recognizing digits via intrinsic robot tactile sensing

    Teresa Sinico, Giovanni Boschetti, and Pedro Neto. Enhancing physical human-robot interaction: Recognizing digits via intrinsic robot tactile sensing. arXiv preprint arXiv:2504.00167,

    arXiv

  3. [18]

    Octo: An open-source generalist robot policy

    Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213,

    arXiv

  4. [19]

    Vlm see, robot do: Human demo video to robot action plan via vision language model

    Beichen Wang, Juexiao Zhang, Shuwen Dong, Irving Fang, and Chen Feng. Vlm see, robot do: Human demo video to robot action plan via vision language model. arXiv preprint arXiv:2410.08792,

  5. [20]

    Gelsight wedge: Measuring high-resolution 3d contact geometry with a compact robot finger

    Shaoxiong Wang, Yu She, Branden Romero, and Edward Adelson. Gelsight wedge: Measuring high-resolution 3d contact geometry with a compact robot finger. In 2021 IEEE International Conference on Robotics and Automation (ICRA) , pp. 6468–6475. IEEE,

    2021

  6. [23]

    Guiding long-horizon task and motion planning with vision language models

    12 Zhutian Yang, Caelan Reed Garrett, Dieter Fox, Tom ´as Lozano-P´erez, and Leslie Pack Kaelbling. Guiding long-horizon task and motion planning with vision language models. In2nd CoRL Work- shop on Learning Effective Abstractions for Planning . Jiawen Yu, Hairuo Liu, Qiaojun...

  7. [24]

    Mimictouch: Leveraging multi-modal human tactile demonstrations for contact-rich manipulation

    Kelin Yu, Yunhai Han, Qixian Wang, Vaibhav Saxena, Danfei Xu, and Ye Zhao. Mimictouch: Leveraging multi-modal human tactile demonstrations for contact-rich manipulation. In 8th An- nual Conference on Robot Learning . Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqi...

    arXiv

  8. [1981]

    A generalist agent

    Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio G ´omez Colmenarejo, Alexander Novikov, Gabriel Barth-maron, Mai Gim ´enez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, et al. A generalist agent. Transactions on Machine Learning Research. Gerrit Schoettler, Ashvin Nai...

    2020

  9. [1985]

    Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning

    Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024 . Binghao Huang, Yixuan Wang, Xinyi Yang, ...

    2024

  10. [1999]

    Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation

    Han Xue, Jieji Ren, Wendi Chen, Gu Zhang, Yuan Fang, Guoying Gu, Huazhe Xu, and Cewu Lu. Reactive diffusion policy: Slow-fast visual-tactile policy learning for contact-rich manipulation. arXiv preprint arXiv:2503.02881,

    arXiv

  11. [2018]

    Measuring and im- proving chain-of-thought reasoning in vision-language models

    Yangyi Chen, Karan Sikka, Michael Cogswell, Heng Ji, and Ajay Divakaran. Measuring and im- proving chain-of-thought reasoning in vision-language models. In Proceedings of the 2024 Con- ference of the North American Chapter of the Association for Computational Linguistics: Huma...

    2024

  12. [2021]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv...

    2010 arXiv

  13. [2022]

    Dexvla: Vision-language model with plug-in diffusion expert for general robot control

    Junjie Wen, Yichen Zhu, Jinming Li, Zhibin Tang, Chaomin Shen, and Feifei Feng. Dexvla: Vision-language model with plug-in diffusion expert for general robot control. arXiv preprint arXiv:2502.05855,

    arXiv

  14. [2023]

    Tla: Tactile-language-action model for contact-rich manipulation

    Peng Hao, Chaofan Zhang, Dingzhe Li, Xiaoge Cao, Xiaoshuai Hao, Shaowei Cui, and Shuo Wang. Tla: Tactile-language-action model for contact-rich manipulation. arXiv preprint arXiv:2503.08548,

    arXiv

  15. [2024]

    π0: A vision-language-action flow model for general robot control

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. π0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164,

    arXiv

  16. [2025]

    Openvla: An open-source vision-language-action model

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P Foster, Pannag R Sanketi, Quan Vuong, et al. Openvla: An open-source vision-language-action model. In 8th Annual Conference on Robot Learning . Fanqi Lin, Ruiqia...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.