Pith. sign in

REVIEW 5 major objections 9 minor 97 references

Predicting future touch, not reacting to current touch, lets a robot policy master contact-rich manipulation at scale.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 12:38 UTC pith:PRZHW2VF

load-bearing objection Solid systems package on predictive tactile VLAs with real margins; the causal story for “predictive latent” vs other confounds is thinner than the tables. the 5 major comments →

arxiv 2607.23782 v1 pith:PRZHW2VF submitted 2026-07-26 cs.RO

N₀-VTLA: Scaling Vision-Tactile-Language-Action Model with Latent Tactile Tokens

NeoteAI Team , Fudan TEAI Team This is my paper
classification cs.RO
keywords vision-tactile-language-actionlatent tactile tokenscontact-rich manipulationflow matchingoffline reinforcement learningadvantage conditioningvisuo-tactile pretrainingrobot foundation model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most general robot policies see cameras and language but not touch, so they struggle when contact decides success—insertions, grip force, deformable objects. This paper argues the right fix is not to dump current tactile images into the vision-language stream, but to train a small predictor that estimates how contact will change over the next action chunk and to condition the action generator on that latent forecast. The model is pretrained on a large multi-platform visuo-tactile corpus, the tactile pathway is brought online in three stages so the pretrained backbone stays stable, and a second method (ALTER) turns stored deployment rollouts into stage-relative advantage labels for offline improvement. On contact-heavy real and simulated benchmarks the approach beats strong vision-only and specialist baselines by large margins, and with ALTER three long-horizon real tasks reach high success. A sympathetic reader cares because the work claims a concrete inductive bias and training recipe that make high-resolution touch usable inside modern language-conditioned robot policies without rebuilding them from scratch.

Core claim

N0-VTLA shows that a vision-language-action backbone can be turned into a strong contact-rich controller by pretraining on large-scale visuo-tactile data and conditioning the action expert on predicted latent tactile tokens—estimates of net contact change over the coming action horizon—rather than on current tactile readings in the vision-language prefix. With that pathway and an offline advantage-labeling method (ALTER), the system wins all nine NeoReal real-robot tasks, reaches 63.8% mean success on a twenty-task simulation suite versus 44.0% for the strongest baseline, and attains 75–95% success on three long-horizon real tasks after offline improvement.

What carries the argument

Latent tactile tokens z: a lightweight predictor reads current contact-difference tokens plus the vision-language context and emits a compact latent that matches the tactile change over the next action chunk; the flow-matching action expert is conditioned on z, while raw touch never enters the vision-language prefix. A three-stage recipe (ground predictor, align expert with prefix masked, then end-to-end) and ALTER (pairwise progress model → stage-relative binary advantage tags on a fixed deployment corpus) carry the training argument.

Load-bearing premise

The load-bearing premise is that a frozen visual encoder on per-episode contact-difference gel images, distilled into a future-horizon latent and refined mainly by the action loss, is a stable enough touch representation that predictive conditioning—not data scale or protocol quirks alone—explains the gains on contact-rich control.

What would settle it

Run the same backbone, data, and evaluation with current-touch injection or free latents (no Stage-1 future-tactile target) versus the predictive latent pathway: if predictive z does not keep a clear success-rate edge on insertion, grip-force, and dual-arm contact tasks, and if Stage-1 retrieval and touch-sensitivity probes collapse, the central design claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Contact-rich skills such as outlet plugging, board insertion, and compliant bottle standing become closed-loop retry and force-regulation behaviors rather than one-shot visual commitments.
  • Large multi-platform visuo-tactile pretraining plus a canonical action container lets one tactile-aware policy cover single- and dual-arm embodiments.
  • Stored deployment data—including drops and human corrections—can raise long-horizon real-task success into the 75–95% range without further environment interaction via advantage-conditioned offline training.
  • A randomly initialized tactile pathway can be grafted onto a pretrained vision-language-action model without destabilizing it if grounding and interface alignment precede joint training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If predictive latents are the right bias, many reactive tactile-VLA designs may be leaving anticipatory pre-contact control on the table even when they already have gel sensors.
  • ALTER’s tactile-grounded drop and stage signals suggest offline robot RL may improve faster when progress labels are tied to contact events than when they rely only on uniform time or sparse final success.
  • The free-latent ablation space the paper flags implies a research fork: how little future-tactile supervision is enough before action gradients alone keep z from becoming a vision shortcut.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 9 minor

Summary. The manuscript presents N0-VTLA, a vision–tactile–language–action model built on a PaliGemma/π0.5-style flow-matching backbone with a novel tactile integration: a frozen DINOv2 encoder turns per-episode contact-difference gel images into tokens, and a lightweight predictor distills them—conditioned on the vision–language prefix—into latent tactile tokens z trained (Stage 1) to match a future-tactile target z* over the coming action chunk (Eq. 2). The action expert is conditioned on z only; tactile never enters the VL prefix. A three-stage recipe grounds the predictor, aligns the latent-to-expert interface under a masked prefix, then trains end to end. A second contribution, ALTER, converts stored deployment data into binary advantage conditions via a pairwise progress model supervised by duration-calibrated stage progress, tactile-detected drop events, and HIL corrections. Empirically, the system wins all nine NeoReal real-robot tasks against internally reproduced ACT and π0.5, reaches 83.1% on UniVTAC (vs 67.1% for the strongest external baseline) and 50.8% on NeoSim, and reaches 75–95% with ALTER on three long-horizon real tasks. Representation analyses show 92.3% top-1 retrieval of future-tactile targets (vs 57% for a current-touch control) and a counterfactual probe showing touch-dependent action divergence at contact.

Significance. If the results hold, this is a substantial contribution: to my knowledge it is the first VLA policy pretrained on tactile data at scale, with a large multi-platform visuo-tactile corpus (NeoData), a canonical cross-embodiment action schema, and publicly stated code/website assets. The empirical package is unusually broad for a tactile-manipulation paper: real-robot evaluation on nine contact-rich tasks, a twenty-task simulation suite with several strong recent external baselines, a progress-rubric scoring scheme with partial credit, throughput numbers for ALTER policies, and direct representation diagnostics (held-out retrieval, a current-touch control, perturbation probes, and a same-noise counterfactual). The "predict, don't react" latent-token design is a genuinely differentiated point in the design space relative to prefix-concatenation and action-path-injection approaches, and ALTER's use of tactile events and HIL logs to supervise an offline progress model is a clean, reusable idea. The contact-aware retry behavior on insertion tasks (§6, Socket Plugging 85% vs 60%) is the kind of qualitative result the field needs.

major comments (5)
  1. [§5 Experiments / §2.2] §5.1–5.3 vs §2.2: the central design claim—conditioning the action expert on *predicted future-contact latents* rather than on current tactile tokens (in the prefix or the action path)—is never ablated in closed loop. The no-tactile control (§5.1, 'tactile flag disabled') isolates 'some tactile pathway,' not the predictive-latent placement. VITaL, ACT+UniVTAC, and Tactile-VLA-style baselines are external systems differing in backbone, data, and training, so they cannot attribute the margin to placement. The paper itself names the two competing placements (§1) as the state of the art it rejects; a controlled internal comparison—same backbone, same NeoData pretraining, same three-stage budget, but with (a) current-touch tokens g prepended to the action suffix and (b) g concatenated into the VL prefix—on a subset of NeoReal/NeoSim tasks is the load-bearing experiment. Without it, the margin
  2. [§4.2 Three-Stage Latent-Tactile Training] The free-latent variant is defined in both §2.2 and §4.2 ('collapses all three stages into a single joint stage') and the three-stage recipe is justified by the prediction that, without Stage 1 supervision, 'nothing stops the bottleneck from degenerating into a copy of the prefix or being bypassed altogether.' No result is reported for this variant, and the Stage 1/2/3 ablations are argued verbally. Since the staging curriculum is one of the paper's three named contributions and the only quantitative evidence offered for it is the retrieval metric (a Stage-1 diagnostic, i.e., the objective itself), at minimum a closed-loop comparison of free-latent vs three-stage on a few tasks—and ideally a per-stage ablation—should be included. Otherwise §4.2's 'Why three stages?' paragraph is an untested hypothesis presented as design rationale.
  3. [§5.5 Representation Analyses] §5.5, perturbation probe: the tactile-to-vision sensitivity ratio falls from 4.3 after Stage 1 to ≈1.4 after end-to-end joint training. The text claims the perturbation probe 'confirms [grounding] is retained rather than rerouted,' but a threefold drop in relative tactile sensitivity after Stage 3—when the InfoNCE/Lrec losses are removed (§4.2) and the direct prefix-to-expert path is restored—is equally consistent with partial rerouting toward the restored prefix pathway. The same-noise counterfactual in Fig. 10 mitigates this for one episode, but the claim that Stage 3 'refines the grounded pathway' needs either (a) the probe statistics over many episodes and tasks, or (b) a control policy trained without Stage 1/2 whose sensitivity ratio is shown to be worse. As written, the sentence 'The predictor keeps its grounding through the action gradient alone' (§4.2) overstates what the eviden
  4. [§5.1 Evaluation Protocol / §5.2] §5.1/§5.2 and Table 2: evaluation reproducibility rests on internal assets—NeoReal rubrics ('full set in the companion data report'), internally reproduced ACT and π0.5 baselines 'under one aligned protocol,' and per-task trial budgets that are never stated. For NeoReal, success rates like 85% vs 60% (Socket Plugging) imply trial counts of perhaps 20–40; without reporting N per task, confidence intervals, and the fine-tuning budget given to each baseline (demonstration counts, compute, tuning parity), the margins cannot be assessed for statistical robustness. Please report per-task trial counts, the demonstration budget and adaptation recipe for each baseline, and ideally binomial confidence intervals or a paired significance test per task. Similarly, NeoSim trains each policy on 100 demonstrations—please state whether baselines received identical data and adaptation compute.
  5. [§6 Findings, Fig. 10] Fig. 10: the mechanism claim 'touch changes the action exactly when contact decides the outcome' rests on a single probed episode at three selected moments. Because this figure is the only evidence that the latent pathway causally shapes closed-loop action selection (as opposed to being predictive in representation space), it should be aggregated: report the path-divergence statistic over a held-out set of episodes, aligned to automatically detected contact events, with a null distribution (e.g., divergence under touch-swap at non-contact moments). One episode, however clean the same-noise construction, does not support the general claim in §6.
minor comments (9)
  1. [§5.5] §5.5: the retrieval evaluation uses 378 held-out queries against pools of ~32 (and 128) candidates. These pool sizes are small relative to the diversity of future-tactile targets in NeoData; please clarify how pools are constructed (same task? same platform? random across corpus?) and report results at larger pools or full-corpus retrieval.
  2. [§2.2, Eqs. (1)–(2)] Eq. (2): z* averages encoder outputs across the n active views into R^{10×d}, while g (Eq. 1) concatenates them into R^{10n×d}. The asymmetry between the predictor's input granularity and its target granularity is worth one sentence of justification—does per-view supervision (10n tokens) work worse, or is view-averaging chosen for the action-chunk-level interpretation?
  3. [§4.4 / Appendix F] §4.4: ALTER's stage templates are generated by Gemini-3.5-Flash with one human review per task. Please quantify annotation reliability: inter-rater or VLM-vs-human agreement on interval-to-stage assignment for a sample of demonstrations, and sensitivity of final policy success to the within-stage positive quantile ρ = 0.3, the event margin m = 0.02, and the 0.3 tag-dropout probability.
  4. [§5.2, Fig. 7] Fig. 7: the upper panel is labeled 'Simulation success rate' but reports real-robot (NeoReal) results—presumably a copy-paste error. Also, per-task y-axis ranges vary between panels (e.g., 0–45, 0–85, 0–25), which visually compresses differences; a common scale or explicit warning is advisable.
  5. [Global] Figure captions and headers use a corrupted glyph for the model name (e.g., '𝒩!-VTLA' in Figs. 1 and 2, and 'AL TER' with a spurious space in §4.4 and the contribution list). Please fix throughout.
  6. [§5.3, Table 2] §5.3: on several NeoSim dual-arm tasks N0-VTLA does not beat π0.5 (Insert Screw 24 vs 26, Stack Plates 94 vs 96, Unstack Cup 10 vs 3 is a win, but the dual-arm mean gap is 39.4 vs 34.3—far narrower than the single-arm gap). The text's framing ('halve it to 39.4%') is honest, but §6's claim that the pathway targets 'bimanual regimes' should be tempered given the small dual-arm margin.
  7. [§5.4] §5.4/Fig. 9: please state the trial count behind each success rate (90/75/60/95/80/75%) and whether the deployment corpus used for ALTER training was collected by the same checkpoints being evaluated (on-policy) or earlier iterations; the distinction matters for interpreting the offline-RL claim.
  8. [§7] Related work (§7): given the emphasis on predictive touch, TTP [88] and Imagine2Touch [3] deserve a sentence of explicit functional comparison (what exactly is transferred/conditioned on), since they are the closest prior art to the latent-token idea.
  9. [Appendix B] Appendix B omits all compute details (hardware, counts, wall-clock). For a paper whose first contribution is pretraining at scale, even aggregate figures (GPU-hours for base pretraining and per-stage post-training) would substantially aid reproducibility and let readers weigh the cost of the tactile pathway against its gains.

Circularity Check

0 steps flagged

Empirical VTLA systems paper: headline claims are external rollout success rates, not quantities forced by construction from fitted inputs or self-citation chains.

full rationale

N0-VTLA’s load-bearing claims are closed-loop success/progress on NeoReal and UniVTAC/NeoSim against external task criteria and reproduced baselines, plus ALTER-improved real-robot success. The latent pathway (Eqs. 1–2, Stages 1–3) is an architectural inductive bias trained with InfoNCE/L1 then action loss; the 92.3% Stage-1 retrieval figure is a diagnostic of that auxiliary objective, not a renamed fit sold as an independent physical prediction. ALTER builds pairwise progress labels from demos/events and conditions policy training; deployment always uses the positive tag, but final metrics remain hardware/simulator task completion, not agreement with the progress model. Companion data report [51] and backbone citations (π0.5, PaliGemma, DINOv2) supply data/protocol and inherited components; they do not import a uniqueness theorem or force the reported margins by definition. No self-definitional loop, fitted-parameter-as-prediction, or load-bearing self-citation chain reduces the central empirical claims to their inputs. Evidence gaps (e.g., missing free-latent closed-loop ablations) are evaluation incompleteness, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 3 invented entities

Load-bearing content is methodological and empirical rather than axiomatic physics. The claim rests on standard deep-learning practice, robotics embodiment conventions, and several paper-specific design choices (predictive latent z, frozen DINO tactile encoder, difference images, staged unfreezing, ALTER labeling rules and thresholds). Free parameters are training/labeling hyperparameters and data-scale claims largely externalized to NeoData.

free parameters (5)
  • Action chunk horizon H = 50
    Fixed at 50 steps for targets, flow matching, and ALTER local change; chosen design hyperparameter that defines the prediction horizon.
  • ALTER within-stage positive quantile ρ = 0.3
    Top fraction labeled Advantage:positive for offline policy learning; directly shapes the binary advantage distribution.
  • Stage-1 loss weight λ_rec and InfoNCE setup = λ_rec > 0 (exact value not highlighted)
    Balances contrastive retrieval vs L1 reconstruction of coarse future tactile difference; set during predictor grounding.
  • ALTER event margin m and event sampling mix = m=0.02; ~5% pairs; multiplier 0.1
    Hinge margin and ~5% event-pair mix with 0.1 loss multiplier control how drops/HIL corrections supervise the progress model.
  • Advantage tag dropout at train time = 0.3
    Tag omitted with probability 0.3 so policy also sees untagged prompts.
axioms (6)
  • domain assumption Flow-matching action experts on a PaliGemma/π0.5-style VLA backbone are a valid base for grafting new sensory pathways.
    Base architecture and weights are inherited from π0.5/PaliGemma (§2.1, §5).
  • domain assumption Vision-based gel tactile difference images encoded by frozen DINOv2 plus a linear projection preserve contact information needed for fine manipulation.
    Eq. 1 and encoder choice in §2.2/§5; absolute gel appearance is discarded by baseline subtraction.
  • ad hoc to paper Future net tactile change over the action chunk is a better conditioning target for contact-rich control than current tactile tokens in the VL prefix or action path.
    Core inductive bias argued in ‘Why predict, not react?’ (§2.2); supported by probes but still a design axiom.
  • domain assumption A canonical 32-D cross-embodiment action/state container lets multi-platform NeoData train one policy.
    §3 schema inherited from π0.5; zero-filled unused dims.
  • ad hoc to paper Duration-calibrated stage progress plus local pairwise comparisons yield useful binary advantages for offline policy improvement without online interaction.
    ALTER formulation §4.4; related to RECAP/χ0 but with paper-specific labeling.
  • standard math Standard supervised learning / contrastive InfoNCE / offline imitation-style flow-matching optimization assumptions (i.i.d. batches, demonstration quality after verification).
    Training objectives in §4.1–4.4.
invented entities (3)
  • Latent tactile tokens z (predictive future-contact latents) independent evidence
    purpose: Compact conditioning signal for the action expert estimating net tactile change over the coming chunk, without putting touch in the VL prefix.
    Central architectural object (§2.2, Eqs. 1–2); probed via retrieval and perturbation but defined by this training recipe.
  • ALTER advantage-labeling procedure independent evidence
    purpose: Convert deployment demos, tactile drop events, and HIL corrections into stage-relative binary advantage text conditions for offline policy training.
    Named method §4.4; builds on prior advantage-conditioned VLA work but introduces specific progress/event machinery.
  • NeoData multi-platform visuo-tactile corpus no independent evidence
    purpose: Supply large-scale pretraining contact priors claimed as first VTLA pretraining at scale.
    Dataset is foundational to the scale claim; details largely in companion report [51].

pith-pipeline@v1.2.0-grok45-kimik3 · 31097 in / 4152 out tokens · 75907 ms · 2026-07-30T12:38:31.922368+00:00 · methodology

0 comments
read the original abstract

We present $N_0$-VTLA, a vision-tactile-language-action (VTLA) foundation model capable of (1) fine-grained contact-rich manipulation with tactile perception and tactile-feedback control, and (2) offline policy improvement from stored deployment data. Building on current vision-based backbones, we propose a training recipe for tactile integration consisting of visuo-tactile pre-training, staged tactile-pathway integration, and advantage-conditioned offline policy improvement. During pre-training, the policy learns broad contact priors from NeoData, our large-scale visuo-tactile robot dataset; to our knowledge, $N_0$-VTLA is the first VTLA model pretrained on tactile data at scale. During post-training, we augment the policy with a predictive tactile pathway that distills the contact patterns learned at scale into the fine motion adjustments required by downstream tactile-centric manipulation. For offline policy improvement, we introduce ALTER, an advantage-conditioned offline reinforcement learning method that converts relative progress and trajectory-event comparisons into binary advantage labels for policy training on a fixed deployment corpus, further improving task-specific learning on contact-rich skills such as deformable object manipulation. Across contact-rich benchmarks, $N_0$-VTLA outperforms strong baselines by wide margins: it wins all nine real-robot NeoReal tasks and reaches 63.8% mean success on a twenty-task simulation suite, against 44.0% for the strongest baseline. $N_0$-VTLA policies trained with ALTER reach 75-95% success on three long-horizon real-robot tasks. These results lay a foundation for versatile tactile-driven manipulation policies.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

97 extracted references · 33 linked inside Pith

  1. [1]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. CVPR, 2023

    2023

  2. [2]

    V-JEPA 2: Self-supervised video models enable understanding, prediction and planning

    Mido Assran, Adrien Bardes, David Fan, Quentin Garrido, Russell Howes, Mojtaba Komeili, Matthew Muckley, Ammar Rizvi, Claire Roberts, Koustuv Sinha, et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning. arXiv preprint arXiv:2506.09985 , 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Imagine2touch: Predictive tactile sensing for robotic manipulation using efficient low-dimensional signals

    Abdallah Ayad, Adrian Röfer, Nick Heppert, and Abhinav Valada. Imagine2touch: Predictive tactile sensing for robotic manipulation using efficient low-dimensional signals. arXiv preprint arXiv:2405.01192, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    LeJEPA: Provable and scalable self-supervised learning without the heuristics

    Randall Balestriero and Yann LeCun. LeJEPA: Provable and scalable self-supervised learning without the heuristics. arXiv preprint arXiv:2511.08544 , 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Revisiting feature prediction for learning visual representations from video

    Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual representations from video. TMLR, 2024

    2024

  6. [6]

    PaliGemma: A versatile 3B VLM for transfer

    Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. PaliGemma: A versatile 3B VLM for transfer. arXiv preprint arXiv:2407.07726 , 2024

    Pith/arXiv arXiv 2024

  7. [7]

    VLA-Touch: Enhancing vision-language-action models with dual-level tactile feedback

    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. RA-L, 2026

    2026

  8. [8]

    GR00T N1: An open foundation model for generalist humanoid robots

    Johan Bjorck, Fernando Castañeda, Nikita Cherniadev, Xingye Da, Runyu Ding, Linxi Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, et al. GR00T N1: An open foundation model for generalist humanoid robots. arXiv preprint arXiv:2503.14734 , 2025

    Pith/arXiv arXiv 2025

  9. [9]

    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. RSS, 2025

    2025

  10. [10]

    RT-2: Vision-language-action models transfer web knowledge to robotic control

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, A vinava Dubey, Chelsea Finn, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. CoRL, 2023

    2023

  11. [11]

    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, Alexander Herzog, Jasmine Hsu, et al. RT-1: Robotics transformer for real-world control at scale. RSS, 2023

    2023

  12. [12]

    UniVLA: Learning to act anywhere with task-centric latent actions

    Qingwen Bu, Yanting Yang, Jisong Cai, Shenyuan Gao, Guanghui Ren, Maoqing Yao, Ping Luo, and Hongyang Li. UniVLA: Learning to act anywhere with task-centric latent actions. RSS, 2025

    2025

  13. [13]

    LeRobot: An open-source library for end-to-end robot learning

    Remi Cadene, Simon Aliberts, Francesco Capuano, Michel Aractingi, Adil Zouitine, Pepijn Kooijmans, Jade Choghari, Martino Russi, Caroline Pascal, Steven Palma, et al. LeRobot: An open-source library for end-to-end robot learning. ICLR, 2026

    2026

  14. [14]

    InternVLA-A1: Unifying understanding, generation and action for robotic manipulation

    Junhao Cai, Zetao Cai, Jiafei Cao, Yilun Chen, Zeyu He, Lei Jiang, Hang Li, Hengjie Li, Yang Li, Yufei Liu, et al. InternVLA-A1: Unifying understanding, generation and action for robotic manipulation. arXiv preprint arXiv:2601.02456 , 2026

    arXiv 2026

  15. [15]

    Xiaomi-Robotics-0: An open-sourced vision-language-action model with real-time execution

    Rui Cai, Jun Guo, Xinze He, Piaopiao Jin, Jie Li, Bingxuan Lin, Futeng Liu, Wei Liu, Fei Ma, Kun Ma, et al. Xiaomi-Robotics-0: An open-sourced vision-language-action model with real-time execution. arXiv preprint arXiv:2602.12684 , 2026. 18

    arXiv 2026

  16. [16]

    GR-3 technical report

    Chilam Cheang, Sijin Chen, Zhongren Cui, Yingdong Hu, Liqun Huang, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Xiao Ma, et al. GR-3 technical report. arXiv preprint arXiv:2507.15493 , 2025

    Pith/arXiv arXiv 2025

  17. [17]

    GR-2: A generative video-language-action model with web-scale knowledge for robot manipulation

    Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, et al. GR-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158 , 2024

    Pith/arXiv arXiv 2024

  18. [18]

    UniVTAC: A unified simulation platform for visuo-tactile manipulation data generation, learning, and benchmarking

    Baijun Chen, Weijie Wan, Tianxing Chen, Xianda Guo, Congsheng Xu, Yuanyang Qi, Haojie Zhang, Longyan Wu, Tianling Xu, Zixuan Li, et al. UniVTAC: A unified simulation platform for visuo-tactile manipulation data generation, learning, and benchmarking. arXiv preprint arXiv:2602.10093 , 2026

    arXiv 2026

  19. [19]

    Multi-modal manipulation via multi-modal policy consensus

    Haonan Chen, Jiaming Xu, Hongyu Chen, Kaiwen Hong, Binghao Huang, Chaoqi Liu, Jiayuan Mao, Yunzhu Li, Yilun Du, and Katherine Driggs-Campbell. Multi-modal manipulation via multi-modal policy consensus. ICRA, 2026

    2026

  20. [20]

    SARM: Stage-aware reward modeling for long horizon robot manipulation

    Qianzhong Chen, Justin Yu, Mac Schwager, Pieter Abbeel, Yide Shentu, and Philipp Wu. SARM: Stage-aware reward modeling for long horizon robot manipulation. ICLR, 2026

    2026

  21. [21]

    Visuo-tactile transformers for manipulation

    Yizhou Chen, Andrea Sipos, Mark Van der Merwe, and Nima Fazeli. Visuo-tactile transformers for manipulation. CoRL, 2022

    2022

  22. [22]

    OmniVTLA: Vision-tactile-language-action models with semantic-aligned tactile sensing

    Zhengxue Cheng, Yiqian Zhang, Anni Tang, Keyu Wang, Wenkang Zhang, Haoyu Li, Hengdi Zhang, and Li Song. OmniVTLA: Vision-tactile-language-action models with semantic-aligned tactile sensing. RA-L, 2026

    2026

  23. [23]

    Diffusion policy: Visuomotor policy learning via action diffusion

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. RSS, 2023

    2023

  24. [24]

    Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots

    Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. RSS, 2024

    2024

  25. [25]

    Gemini 3.5 flash

    Google DeepMind. Gemini 3.5 flash. https://deepmind.google/models/gemini/flash/, 2025

    2025

  26. [26]

    Tenenbaum, Dale Schuurmans, and Pieter Abbeel

    Yilun Du, Mengjiao Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Joshua B. Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation. NeurIPS, 2023

    2023

  27. [27]

    LIBERO-Plus: In-depth robustness analysis of vision-language-action models

    Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, Jinlan Fu, Jingjing Gong, and Xipeng Qiu. LIBERO-Plus: In-depth robustness analysis of vision-language-action models. CVPR, 2026

    2026

  28. [28]

    AnyTouch: Learning unified static-dynamic representation across multiple visuo-tactile sensors

    Ruoxuan Feng, Jiangyu Hu, Wenke Xia, Tianci Gao, Ao Shen, Yuhao Sun, Bin Fang, and Di Hu. AnyTouch: Learning unified static-dynamic representation across multiple visuo-tactile sensors. ICLR, 2025

    2025

  29. [29]

    VITaL pretraining: Visuo-tactile pretraining for tactile and non-tactile manipulation policies

    Abraham George, Selam Gano, Pranav Katragadda, and Amir Barati Farimani. VITaL pretraining: Visuo-tactile pretraining for tactile and non-tactile manipulation policies. ICRA, 2025

    2025

  30. [30]

    Hap- ticVLA: Contact-rich manipulation via vision-language-action model without inference-time tactile sens- ing

    Konstantin Gubernatorov, Mikhail Sannikov, Ilya Mikhalchuk, Egor Kuznetsov, Makar Artemov, Ogun- woye Faith Ouwatobi, Marcelino Fernando, Artem Asanov, Ziang Guo, and Dzmitry Tsetserukou. Hap- ticVLA: Contact-rich manipulation via vision-language-action model without inference-time tactile sens- ing. arXiv preprint arXiv:2603.15257 , 2026

    arXiv 2026

  31. [31]

    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. Robot Learning, 2026

    2026

  32. [32]

    MiMo-Embodied: X-embodied foundation model technical report

    Xiaoshuai Hao, Lei Zhou, Zhijian Huang, Zhiwen Hou, Yingbo Tang, Lingfeng Zhang, Guang Li, Zheng Lu, Shuhuai Ren, Xianhui Meng, et al. MiMo-Embodied: X-embodied foundation model technical report. arXiv preprint arXiv:2511.16518 , 2025. 19

    Pith/arXiv arXiv 2025

  33. [33]

    Sparsh: Self- supervised touch representations for vision-based tactile sensing

    Carolina Higuera, Akash Sharma, Chaithanya Krishna Bodduluri, Taosha Fan, Patrick Lancaster, Mri- nal Kalakrishnan, Michael Kaess, Byron Boots, Mike Lambeta, Tingfan Wu, et al. Sparsh: Self- supervised touch representations for vision-based tactile sensing. CoRL, 2024

    2024

  34. [34]

    Visuo-tactile world models

    Carolina Higuera, Sergio Arnaud, Byron Boots, Mustafa Mukadam, Francois Robert Hogan, and Franziska Meier. Visuo-tactile world models. arXiv preprint arXiv:2602.06001 , 2026

    arXiv 2026

  35. [35]

    3D-ViTac: Learning fine-grained manipulation with visuo-tactile sensing

    Binghao Huang, Yixuan Wang, Xinyi Yang, Yiyue Luo, and Yunzhu Li. 3D-ViTac: Learning fine-grained manipulation with visuo-tactile sensing. CoRL, 2024

    2024

  36. [36]

    Tactile-VLA: Un- locking vision-language-action model’s physical knowledge for tactile generalization

    Jialei Huang, Shuo Wang, Fanqi Lin, Yihang Hu, Chuan Wen, and Yang Gao. Tactile-VLA: Un- locking vision-language-action model’s physical knowledge for tactile generalization. arXiv preprint arXiv:2507.09160, 2025

    Pith/arXiv arXiv 2025

  37. [37]

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

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. OpenVLA: An open-source vision-language- action model. CoRL, 2024

    2024

  38. [38]

    ProgVLA: Progress-aware robot manipulation skill learning

    Seungsu Kim, Jinyoung Choi, Seungmin Baek, and Jean-Michel Renders. ProgVLA: Progress-aware robot manipulation skill learning. arXiv preprint arXiv:2605.28231 , 2026

    Pith/arXiv arXiv 2026

  39. [39]

    DIGIT: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation

    Mike Lambeta, Po-Wei Chou, Stephen Tian, Brian Yang, Benjamin Maloon, Victoria Rose Most, Dave Stroud, Raymond Santos, Ahmad Byagowi, Gregg Kammerer, et al. DIGIT: A novel design for a low-cost compact high-resolution tactile sensor with application to in-hand manipulation. RA-L, 2020

    2020

  40. [40]

    CronusVLA: Towards efficient and robust manipulation via multi-frame vision-language-action modeling

    Hao Li, Shuai Yang, Yilun Chen, Xinyi Chen, Xiaoda Yang, Yang Tian, Hanqing Wang, Tai Wang, Dahua Lin, Feng Zhao, et al. CronusVLA: Towards efficient and robust manipulation via multi-frame vision-language-action modeling. AAAI, 2026

    2026

  41. [41]

    Xiaomi-Robotics-U0: Unified embodied synthesis with world foundation model

    Xinghang Li, Jun Guo, Qiwei Li, Long Qian, Hang Lai, Yueze Wang, Hongyu Yan, Jiahang Cao, Xi Chen, Jingen Qu, et al. Xiaomi-Robotics-U0: Unified embodied synthesis with world foundation model. arXiv preprint arXiv:2607.11643 , 2026

    Pith/arXiv arXiv 2026

  42. [42]

    GR-RL: Going dexterous and precise for long-horizon robotic manipulation

    Yunfei Li, Xiao Ma, Jiafeng Xu, Yu Cui, Zhongren Cui, Zhigang Han, Liqun Huang, Tao Kong, Yuxiao Liu, Hao Niu, et al. GR-RL: Going dexterous and precise for long-horizon robotic manipulation. arXiv preprint arXiv:2512.01801 , 2025

    arXiv 2025

  43. [43]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. ICLR, 2023

    2023

  44. [44]

    ViTaMIn: Learning contact-rich tasks through robot-free visuo-tactile manipulation interface

    Fangchen Liu, Chuanyu Li, Yihua Qin, Jing Xu, Pieter Abbeel, and Rui Chen. ViTaMIn: Learning contact-rich tasks through robot-free visuo-tactile manipulation interface. arXiv preprint arXiv:2504.06156, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Dream-Tac: A unified tactile world action model for contact-rich robot manipulation

    Yunfan Lou, Yifan Ye, Yankai Fu, Jun Cen, Xiaowei Chi, Yaoxu Lyu, Peidong Jia, Sirui Han, Zhihe Lu, and Shanghang Zhang. Dream-Tac: A unified tactile world action model for contact-rich robot manipulation. arXiv preprint arXiv:2606.08737 , 2026

    Pith/arXiv arXiv 2026

  46. [46]

    VLA-RL: Towards masterful and general robotic manipulation with scalable reinforcement learning

    Guanxing Lu, Wenkai Guo, Chubin Zhang, Yuheng Zhou, Haonan Jiang, Zifeng Gao, Yansong Tang, and Ziwei Wang. VLA-RL: Towards masterful and general robotic manipulation with scalable reinforcement learning. arXiv preprint arXiv:2505.18719 , 2025

    Pith/arXiv arXiv 2025

  47. [47]

    InternVLA-A1.5: Unifying understanding, latent foresight, and action for compositional generalization

    Haoxiang Ma, Junhao Cai, Xiaoxu Xu, Hao Li, Yuyin Yang, Yang Tian, Jiafei Cao, Hongrui Zhu, Zherui Qiu, Zhaxizhuoma, et al. InternVLA-A1.5: Unifying understanding, latent foresight, and action for compositional generalization. arXiv preprint arXiv:2607.04988 , 2026

    Pith/arXiv arXiv 2026

  48. [48]

    Vision language models are in-context value learners

    Yecheng Jason Ma, Joey Hejna, Ayzaan Wahid, Chuyuan Fu, Dhruv Shah, Jacky Liang, Zhuo Xu, Sean Kirmani, Peng Xu, Danny Driess, et al. Vision language models are in-context value learners. ICLR, 2025. 20

    2025

  49. [49]

    LeWorldModel: Stable end-to-end joint-embedding predictive architecture from pixels

    Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. LeWorldModel: Stable end-to-end joint-embedding predictive architecture from pixels. arXiv preprint arXiv:2603.19312 , 2026

    Pith/arXiv arXiv 2026

  50. [50]

    ARM: Advantage reward modeling for long-horizon manipulation

    Yiming Mao, Zixi Yu, Weixin Mao, Yinhao Li, Qirui Hu, Zihan Lan, Minzhao Zhu, and Hua Chen. ARM: Advantage reward modeling for long-horizon manipulation. arXiv preprint arXiv:2604.03037 , 2026

    Pith/arXiv arXiv 2026

  51. [51]

    N0-foundation: Towards the age of tactile intelligence

    NeoteAI Team and Fudan TEAI Team. N0-foundation: Towards the age of tactile intelligence. https: //research.neoteai.com/n0-foundation/, 2026

    2026

  52. [52]

    T-Rex: Tactile-reactive dexterous manipulation

    Dantong Niu, Zhuoyang Liu, Zekai Wang, Boning Shao, Zhao-Heng Yin, Anirudh Pai, Yuvan Sharma, Stefano Saravalle, Ruijie Zheng, Jing Wang, et al. T-Rex: Tactile-reactive dexterous manipulation. arXiv preprint arXiv:2606.17055 , 2026

    Pith/arXiv arXiv 2026

  53. [53]

    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, Jianlan Luo, et al. Octo: An open-source generalist robot policy. RSS, 2024

    2024

  54. [54]

    Open X-embodiment: Robotic learning datasets and RT-X models

    Open X-Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Mad- dukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Man- dlekar, et al. Open X-embodiment: Robotic learning datasets and RT-X models. ICRA, 2024

    2024

  55. [55]

    DINOv2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. TMLR, 2024

    2024

  56. [56]

    SOP: A scalable online post-training system for vision-language-action models

    Mingjie Pan, Siyuan Feng, Qinglin Zhang, Xinchen Li, Jianheng Song, Chendi Qu, Yi Wang, Chuankang Li, Ziyu Xiong, Zhi Chen, et al. SOP: A scalable online post-training system for vision-language-action models. arXiv preprint arXiv:2601.03044 , 2026

    arXiv 2026

  57. [57]

    π∗ 0.6: a VLA that learns from experience

    Physical Intelligence, Ali Amin, Raichelle Aniceto, Ashwin Balakrishna, Kevin Black, Ken Conley, Grace Connors, James Darpinian, Karan Dhabalia, Jared DiCarlo, Danny Driess, et al. π∗ 0.6: a VLA that learns from experience. arXiv preprint arXiv:2511.14759 , 2025

    Pith/arXiv arXiv 2025

  58. [58]

    Galliker, et al

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Manuel Y. Galliker, et al. π0.5: a vision- language-action model with open-world generalization. CoRL, 2025

    2025

  59. [59]

    Generic event boundary detection: A benchmark for event segmentation

    Mike Zheng Shou, Stan Weixian Lei, Weiyao Wang, Deepti Ghadiyaram, and Matt Feiszli. Generic event boundary detection: A benchmark for event segmentation. ICCV, 2021

    2021

  60. [60]

    SmolVLA: A vision-language- action model for affordable and efficient robotics

    Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. SmolVLA: A vision-language- action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844 , 2025

    Pith/arXiv arXiv 2025

  61. [61]

    StarVLA: A lego-like codebase for vision-language-action model developing

    StarVLA Community. StarVLA: A lego-like codebase for vision-language-action model developing. arXiv preprint arXiv:2604.05014 , 2026

    Pith/arXiv arXiv 2026

  62. [62]

    World guidance: World modeling in condition space for action generation

    Yue Su, Sijin Chen, Haixin Shi, Mingyu Liu, Zhengshen Zhang, Ningyuan Huang, Weiheng Zhong, Zhengbang Zhu, Yuxiao Liu, and Xihui Liu. World guidance: World modeling in condition space for action generation. arXiv preprint arXiv:2602.22010 , 2026

    arXiv 2026

  63. [63]

    NeuralFeels with neural fields: Visuotac- tile perception for in-hand manipulation

    Sudharshan Suresh, Haozhi Qi, Tingfan Wu, Taosha Fan, Luis Pineda, Mike Lambeta, Jitendra Malik, Mrinal Kalakrishnan, Roberto Calandra, Michael Kaess, et al. NeuralFeels with neural fields: Visuotac- tile perception for in-hand manipulation. Sci. Robot., 2024

    2024

  64. [64]

    Robo-Dopamine: General process reward modeling for high- precision robotic manipulation

    Huajie Tan, Sixiang Chen, Yijie Xu, Zixiao Wang, Yuheng Ji, Cheng Chi, Yaoxu Lyu, Zhongxia Zhao, Xiansheng Chen, Peterson Co, et al. Robo-Dopamine: General process reward modeling for high- precision robotic manipulation. CVPR, 2026. 21

    2026

  65. [65]

    VT-W AM: Visual-tactile world action model for contact-rich manipulation

    Shuai Tian, Yupeng Zheng, Yuhang Zheng, Songen Gu, Yujie Zang, Yuxing Qin, Weize Li, Haoran Li, Wenchao Ding, and Dongbin Zhao. VT-W AM: Visual-tactile world action model for contact-rich manipulation. arXiv preprint arXiv:2607.02503 , 2026

    Pith/arXiv arXiv 2026

  66. [66]

    Representation learning with contrastive predictive coding

    Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 , 2018

    Pith/arXiv arXiv 2018

  67. [67]

    Qwen-VLA: Unifying vision-language-action modeling across tasks, environments, and robot embodiments

    Qiuyue Wang, Mingsheng Li, Jian Guan, Jinhui Ye, Sicheng Xie, Yitao Liu, Junhao Chen, Zhixuan Liang, Jie Zhang, Xintong Hu, et al. Qwen-VLA: Unifying vision-language-action modeling across tasks, environments, and robot embodiments. arXiv preprint arXiv:2605.30280 , 2026

    Pith/arXiv arXiv 2026

  68. [68]

    Learning while deploying: Fleet-scale reinforcement learning for generalist robot policies

    Yi Wang, Xinchen Li, Pengwei Xie, Pu Yang, Buqing Nie, Yunuo Cai, Qinglin Zhang, Chendi Qu, Jeffrey Wu, Jianheng Song, et al. Learning while deploying: Fleet-scale reinforcement learning for generalist robot policies. arXiv preprint arXiv:2605.00416 , 2026

    Pith/arXiv arXiv 2026

  69. [69]

    Unleashing large-scale video generative pre-training for visual robot manipulation

    Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. ICLR, 2024

    2024

  70. [70]

    FreeTacMan: Robot-free visuo-tactile data collection system for contact-rich manipulation

    Longyan Wu, Checheng Yu, Jieji Ren, Li Chen, Yufei Jiang, Ran Huang, Guoying Gu, and Hongyang Li. FreeTacMan: Robot-free visuo-tactile data collection system for contact-rich manipulation. ICRA, 2026

    2026

  71. [71]

    Tactile-W AM: Touch-aware world action model with tactile asymmetric attention

    Siyu Wu, Linjing You, Junjie Zhu, Yaozu Liu, Changhao Zhang, Jian Liu, Weiqiang Wang, Qi Li, Jituo Li, and Hengshuang Zhao. Tactile-W AM: Touch-aware world action model with tactile asymmetric attention. arXiv preprint arXiv:2606.26663 , 2026

    Pith/arXiv arXiv 2026

  72. [72]

    A pragmatic VLA foundation model

    Wei Wu, Fan Lu, Yunnan Wang, Shuai Yang, Shi Liu, Fangjing Wang, Qian Zhu, He Sun, Yong Wang, Shuailei Ma, et al. A pragmatic VLA foundation model. arXiv preprint arXiv:2601.18692 , 2026

    Pith/arXiv arXiv 2026

  73. [73]

    Robo-ValueRL: Reliable value estimation for offline-to-online reinforcement learning

    Wenke Xia, Pei Ren, Wenbo Yu, Yizhuo Zhang, Jifan Li, Yixue Zhang, Yinuo Zhao, Qingyang Gao, Jian- long Fu, Jian Tang, et al. Robo-ValueRL: Reliable value estimation for offline-to-online reinforcement learning. arXiv preprint arXiv:2607.09866 , 2026

    Pith/arXiv arXiv 2026

  74. [74]

    Xiaomi-Robotics-1: Scaling vision-language-action models with over 100K hours of real-world trajectories

    Xiaomi Robotics Team, Jun Guo, Piaopiao Jin, Jason Li, Peiyan Li, Yingyan Li, Futeng Liu, Wanli Peng, Optimus Qin, Yifei Su, Nan Sun, et al. Xiaomi-Robotics-1: Scaling vision-language-action models with over 100K hours of real-world trajectories. arXiv preprint arXiv:2607.15330 , 2026

    Pith/arXiv arXiv 2026

  75. [75]

    Seeing touch from motion: A unified modality-aware visuo-tactile policy with tactile motion correlation

    Shengqi Xu, Guojin Zhong, Yang Liu, Fanjie Wang, Hu Luo, Hanyu Zhou, Weiyao Zhang, Ziyi Ye, Zuxuan Wu, and Yu-Gang Jiang. Seeing touch from motion: A unified modality-aware visuo-tactile policy with tactile motion correlation. ECCV, 2026

    2026

  76. [76]

    exUMI: Extensible robot teaching system with action-aware task-agnostic tactile representation

    Yue Xu, Litao Wei, Pengyu An, Qingyu Zhang, and Yong-Lu Li. exUMI: Extensible robot teaching system with action-aware task-agnostic tactile representation. CoRL, 2025

    2025

  77. [77]

    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. RSS, 2025

    2025

  78. [78]

    ProgressVLA: Progress-guided diffusion policy for vision-language robotic manipulation

    Hongyu Yan, Qiwei Li, Jiaolong Yang, and Yadong Mu. ProgressVLA: Progress-guided diffusion policy for vision-language robotic manipulation. arXiv preprint arXiv:2603.27670 , 2026

    arXiv 2026

  79. [79]

    Touch and go: Learning from human-collected vision and touch

    Fengyu Yang, Chenyang Ma, Jiacheng Zhang, Jing Zhu, Wenzhen Yuan, and Andrew Owens. Touch and go: Learning from human-collected vision and touch. NeurIPS, 2022

    2022

  80. [80]

    Binding touch to everything: Learning unified multimodal tactile representations

    Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, et al. Binding touch to everything: Learning unified multimodal tactile representations. CVPR, 2024. 22

    2024

Showing first 80 references.