Robotics · arXiv:2609.20761 v2 · submitted 17 Sep 2026
Agile-WAM: An Agile Tactile World Action Model for Contact-Rich Robot Control
The paper says
“World Action Models (WAMs) advance beyond conventional visuomotor policies by jointly predicting future world states and robot actions, enabling the policy to learn phys- ical dynamics that support effective control.”
more of the abstract
However, recent tactile WAMs often rely on large-scale pretrained generative backbones to capture contact-rich physical dynamics, which limit their inference efficiency and flexible deployment. In this paper, we present Agile-WAM, an agile tactile World Action Model for contact-rich robot control. Agile-WAM encodes visual and tactile observations into a shared latent that serves as the source of a direct vision-tactile-to-action flow-matching process, which can jointly generate latent representations of action chunks and future visual/tactile latents. A key observation is that vision and tactile signals evolve at inherently different timescales: adjacent visual frames are often highly similar, whereas tactile signals can change abruptly upon contact. We therefore introduce multi-horizon multimodal prediction in Agile-WAM, which provides supervision for visual latent at a larger temporal offset while predicting the tactile latent in the next frame to capture fine-grained contact dynamics. Across nine simulated and five real-world contact-rich ma- nipulation tasks, Agile-WAM demonstrates strong and robust performance, outperforming the strongest baseline in success rate while maintaining low inference latency. In particular, in five real-world experiments, Agile-WAM yields a relative gain of 29.4% in overall success rates while achieving inference latency of 11.9 ms. These results demonstrate that multimodal WAM can be achieved with an agile architecture suitable for precise and high-frequency robot control. More details are available on our project page.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 20 Sep 2026 · of version 1 of the paper; version 2 is newer · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
The passes disagreed. Reader: holds up. Skeptic: holds up only with conditions. Referee: major revision.
Solid architecture and simulated experiments, but the headline real-world success claim needs statistical backing. 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 references say.
The paper says, at Table II, Section IV-B1
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
Real-world success rates are reported from only 20 episodes per task per policy, with no confidence intervals, error bars, or statistical significance tests. The claimed 29.4% relative overall improvement is not supported by the presented data. For example, on Gear Assembly (80% vs 55%) and Power Plug Insertion (65% vs 35%), the differences are not statistically significant at α=0.05 (Fisher's exact test gives p≈0.09 and p≈0.057 respectively, by my calculation).
more of objection 1
The authors must provide statistical measures (e.g., confidence intervals, p-values) or substantially increase the number of trials to support the central performance claim. At minimum, the limitations of the real-world evaluation should be explicitly discussed and the headline claim tempered.
The paper says, at Section III-C, Fig. 5(c)
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The multi-horizon prediction uses a fixed visual horizon h_vis = H = 16 (the executed action length) and a fixed tactile horizon h_tac = 1. The ablation in Fig. 5(c) shows sensitivity to horizon choices, but the paper does not justify why this specific pair is universally optimal across tasks with very different temporal dynamics.
more of objection 2
The authors should either provide a sensitivity analysis across a range of horizon choices on multiple tasks, or discuss the task-dependent nature of this hyperparameter and provide guidance for selecting it in new applications.
The paper says, at Section IV-A (Training and Evaluation), Table II
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The real-world experiments use a single run of 20 episodes per condition. In contrast, simulation experiments use 50 rollouts per seed with 3 seeds. The discrepancy in statistical rigor between simulation and real-world evaluation undermines confidence in the real-world results. The authors should either match the simulation protocol (multiple seeds, more episodes) or clearly state the real-world results as preliminary and provide appropriate caveats.
5 minor objections
The paper says, at Table I
Pith says: minor objection 1 of 5
The standard deviations in Table I appear to be computed over three seeds. Reporting these as '±' values is appropriate, but the number of seeds (3) should be stated explicitly in the table caption.
The paper says, at Section IV-B1, last paragraph
Pith says: minor objection 2 of 5
The description of corrective behaviors (e.g., sliding the peg along the surface) is qualitative and cannot be verified from the reported aggregate success rates. Consider including a qualitative video or quantitative metrics such as contact force traces to substantiate the claimed recovery mechanism.
show 3 more
The paper says, at Abstract and Section IV-B2
Pith says: minor objection 3 of 5
The abstract states an inference latency of 11.9 ms, but Table III reports 10.35±0.14 ms. Please clarify the source of this discrepancy. Also, the relative control frequency values in Table III are computed relative to VITA; state this explicitly in the table caption.
The paper says, at Section III-C
Pith says: minor objection 4 of 5
The formula for L_vis and L_tac uses squared L2 norm, while L_act and L_AE use L1 norm. The authors should briefly justify this choice (e.g., L2 for latent prediction, L1 for action reconstruction).
The paper says, at General
Pith says: minor objection 5 of 5
Code is not released. Although not required, providing code would significantly enhance reproducibility and impact, especially given the complexity of the architecture.
What would settle it
Re-run all five real-world tasks with at least 50 episodes per policy per task (ideally 100) and compute 95% binomial confidence intervals for each success rate and for the aggregate relative gain. If the 29.4% relative gain is not statistically significant (e.g., confidence intervals overlap or one-sided p > 0.05), the central claim weakens.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 20 Sep 2026
- Reader holds up
- Skeptic partly agrees: says it holds up only with conditions Real-world success rate advantage may not be statistically significant given 20 episodes per condition.
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The main new thing here is multi-horizon multimodal prediction—supervising visual latents at a longer horizon and tactile latents at the next frame to match their different temporal scales. That is a sensible adaptation and the ablation confirms it helps. The paper also shows that a lightweight flow-matching backbone (no cross-attention, no big pretrained generator) can jointly predict actions and future latents at 11.9 ms inference, which is genuinely efficient. The simulated experiments are solid: 9 tasks, 3 seeds, 50 rollouts per evaluation, consistent improvements over baselines. The ablation isolating the benefit of joint action–future modeling (vs. just a larger latent) is cleanly done.
The soft spot is the real-world evaluation. The headline 29.4 % relative gain rests on 20 episodes per condition with no confidence intervals, no significance test, and individual task p-values hovering around 0.06–0.09. The stress-test note is right: this is not robust evidence for the headline claim. The paper explicitly highlights real-world performance as key evidence, so this is a load-bearing weakness. Also, no code release makes it hard to verify the claims.
That said, the simulation evidence is much stronger (50 rollouts × 3 seeds) and shows a consistent, if moderate, advantage. The core architecture is well-motivated and the multi-horizon idea is reusable. I think this paper deserves a serious referee—the idea is sound and the simulated experiments are convincing—but the real-world section needs a statistical overhaul before it can support the paper's strongest claims. A good reviewer will ask for error bars, a proper test (e.g., permutation test across tasks), and ideally code release.
Bottom line: useful for anyone working on tactile policy learning. I'd accept the review, but I'd expect the authors to tighten the statistics before publication.
The strongest objection
The central claim of a 29.4% relative improvement in overall success rates on real-world tasks rests on 20 episodes per task per policy (5 tasks, 3 policies, 300 total). Individual task differences are not statistically significant at α=0.05 (e.g., Agle-WAM vs VITA-VT on Gear Assembly: 80% vs 55%, p≈0.09; Power Plug Insertion: 65% vs 35%, p≈0.057; other tasks show smaller margins). The paper does not report confidence intervals or conduct any statistical test for the overall gain. Without accounting for sampling variability, the headline result may not be robust. The simulation results use 50 rollouts per seed and 3 seeds, but the real-world evaluation uses a single run of 20 trials per setting. This is a load-bearing concern because the paper explicitly highlights real-world performance as its key evidence.
The reader's weakest_assumption focused on temporal horizon generalization, which is a secondary concern. The more fundamental issue is the statistical power of the real-world evaluation supporting the headline claim. The paper reports strong numerical advantages, but with only 20 trials per condition, random variation could explain the observed differences. The reader's ACCEPT verdict did not flag this, but it is a load-bearing concern because the paper emphasizes real-world results. Therefore, I recommend a CONDITIONAL verdict pending additional trials and proper statistical reporting. If the test confirms significance, the original ACCEPT stands; if not, the claim should be downridden.
The referee's report
If the results hold, Agile-WAM addresses an important practical need for efficient tactile world models that can run at high control frequencies without large pretrained backbones. The multi-horizon prediction idea is physically motivated and could influence future multimodal policy architectures. The extensive simulation ablations provide useful insights into the contribution of each design component. However, the significance is tempered by the statistical weakness of the real-world evaluation, which is the primary evidence for the headline claim.
Pith's reading of the paper (a machine's summary, not the authors')
Agile-WAM is an agile tactile World Action Model that jointly predicts action chunks and future visual and tactile latents using a direct vision-tactile-to-action flow-matching process, avoiding large pretrained generative backbones. The key insight is that vision and tactile signals evolve at different timescales: adjacent visual frames are often redundant while tactile signals change abruptly on contact. To address this, the model uses a longer prediction horizon for visual latents (equal to the executed action length) and a next-frame horizon for tactile latents, preserving fine-grained contact dynamics. This design yields a 29.4% relative improvement in overall success rates across five real-world contact-rich manipulation tasks while maintaining an inference latency of 11.9 ms, demonstrating that tactile world action models can be both effective and computationally efficient for high-frequency robot control.
What carries the argument. The central mechanism is a vision-tactile-to-action flow-matching backbone combined with multi-horizon multimodal prediction. The flow-matching backbone directly transports a fused observation latent (encoding current vision, tactile, and proprioception) through a learned velocity field to a target latent that splits into action, future visual, and future tactile components. Multi-horizon prediction then supervises visual latents at a longer horizon (matching the action chunk length) and tactile latents at the next frame, aligning supervision with each modality's temporal scale.
Core claim. The paper introduces Agile-WAM, which uses a fused latent representation of current visual and tactile observations as the source of a flow-matching process that directly generates latent representations of action chunks and future visual and tactile latents, without requiring cross-attention or other conditioning modules. A central design choice is multi-horizon multimodal prediction: visual future latents are supervised at a larger temporal offset (equal to the action horizon), while tactile future latents are predicted at the next time step. This matches the inherent temporal dynamics of each modality—slow-changing vision and fast-changing touch—providing effective supervision for world modeling. Experiments on nine simulated and five real-world tasks show that Agile-WAM outperforms prior vision-only and vision-tactile policies, achieving a 29.4% relative improvement in real-world success rate and inference latency of 11.9 ms, making it suitable for precise, high-frequency robot control.
The premise it rests on. The fixed multi-horizon choice (visual horizon equal to action length, tactile horizon equal to 1) assumes these specific offsets are universally optimal across all contact-rich tasks with different temporal dynamics.
If the paper is right (Pith's inferences, not the authors')
Tactile world action models can be deployed on resource-constrained edge platforms because Agile-WAM achieves 11.9 ms inference latency, potentially enabling high-frequency control without dedicated GPU servers.
Jointly predicting future visual and tactile latents together with actions allows policies to recover from initial misalignment in contact-rich tasks by using touch feedback when vision is occluded.
The multi-horizon prediction principle generalizes to other robotic setups where different sensing modalities (e.g., force, audio, proprioception) evolve at distinct timescales.
Agile-WAM reduces the gap between computationally expensive generative world models and efficient visuomotor policies, making world action modeling practical for real-time manipulation.
The fixed choice of visual horizon (equal to action length) and tactile horizon (next frame) may not be optimal for all tasks; adapting these horizons per task or per episode could further improve performance.
Because the flow-matching backbone is compact and trained from scratch, Agile-WAM could be transferred to new tactile sensor types or visual domains with minimal re-engineering, though this is not tested in the paper.
The multi-horizon idea could be applied to other multi-modal prediction problems outside robotics, such as autonomous driving (e.g., slow visual scene vs. fast LIDAR returns).
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
World Action Models (WAMs) advance beyond conventional visuomotor policies by jointly predicting future world states and robot actions, enabling the policy to learn phys- ical dynamics that support effective control. However, recent tactile WAMs often rely on large-scale pretrained generative backbones to capture contact-rich physical dynamics, which limit their inference efficiency and flexible deployment. In this paper, we present Agile-WAM, an agile tactile World Action Model for contact-rich robot control. Agile-WAM encodes visual and tactile observations into a shared latent that serves as the source of a direct vision-tactile-to-action flow-matching process, which can jointly generate latent representations of action chunks and future visual/tactile latents. A key observation is that vision and tactile signals evolve at inherently different timescales: adjacent visual frames are often highly similar, whereas tactile signals can change abruptly upon contact. We therefore introduce multi-horizon multimodal prediction in Agile-WAM, which provides supervision for visual latent at a larger temporal offset while predicting the tactile latent in the next frame to capture fine-grained contact dynamics. Across nine simulated and five real-world contact-rich ma- nipulation tasks, Agile-WAM demonstrates strong and robust performance, outperforming the strongest baseline in success rate while maintaining low inference latency. In particular, in five real-world experiments, Agile-WAM yields a relative gain of 29.4% in overall success rates while achieving inference latency of 11.9 ms. These results demonstrate that multimodal WAM can be achieved with an agile architecture suitable for precise and high-frequency robot control. More details are available on our project page.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23
The authors' main claim, as the classifier picked it out 0.97
“These results demonstrate that multimodal WAM can be achieved with an agile architecture suitable for precise and high-frequency robot control.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.10 · human participants 0.04 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.24
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 |
|---|---|---|---|---|
| World Action Models (WAMs) advance beyond conventional visuomotor policies by jointly predicting future world states and robot actions, enabling the policy to learn physical dynamics that support effective control. | background 1.00 | 0.05 | 0.04 | asserted 1.00 |
| However, recent tactile WAMs often rely on large-scale pretrained generative backbones to capture contact-rich physical dynamics, which limit their inference efficiency and flexible deployment. | gap 0.53 | 0.06 | 0.02 | asserted 0.97 |
| In this paper, we present \ABBR{}, an agile tactile World Action Model for contact-rich robot control. | contribution 1.00 | 0.85 | 0.06 | asserted 0.96 |
| \ABBR{} encodes visual and tactile observations into a shared latent that serves as the source of a direct vision-tactile-to-action flow-matching process, which can jointly generate latent representations of action chunks and future visual/tactile latents. | method 0.97 | 0.84 | 0.05 | asserted 1.00 |
| A key observation is that vision and tactile signals evolve at inherently different timescales: adjacent visual frames are often highly similar, whereas tactile signals can change abruptly upon contact. | background 0.65 | 0.34 | 0.03 | asserted 0.95 |
| We therefore introduce multi-horizon multimodal prediction in \ABBR{}, which provides supervision for visual latent at a larger temporal offset while predicting the tactile latent in the next frame to capture fine-grained contact dynamics. | method 0.82 | 0.93 | 0.05 | asserted 0.99 |
| Across nine simulated and five real-world contact-rich manipulation tasks, \ABBR{} demonstrates strong and robust performance, outperforming the strongest baseline in success rate while maintaining low inference latency. | result 1.00 | 0.96 | 0.42 | asserted 1.00 |
| In particular, in five real-world experiments, \ABBR{} yields a relative gain of $\textbf{29.4\%}$ in overall success rates while achieving inference latency of $\textbf{11.9 ms}$. | result 1.00 | 0.97 | 0.15 | asserted 1.00 |
| These results demonstrate that multimodal WAM can be achieved with an agile architecture suitable for precise and high-frequency robot control. | interpretation 1.00 | 0.69 | 0.78 | asserted 1.00 |
| More details are available on our project page: https://hanchuzhou.github.io/TARO_project_page/. | future work 0.81 | 0.11 | 0.02 | unclear 0.73 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 7,741 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 (5)
Figures are the authors', reproduced from the paper.
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
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