acceptodds
Under review as a conference paper at ICLR 2027

PhysCanvas: A Vision-Tactile World Model for Robot Policy Evaluation and Learning

Abstract

Evaluating and post-training robot policies through physical interaction is costly and difficult to scale, motivating action-conditioned world models as interactive simulators. However, visually plausible predictions can misrepresent contact dynamics and task outcomes, providing unreliable feedback for policy evaluation and learning. We introduce PhysCanvas, a state-consistent visuo-tactile world model for contact-rich robot manipulation. PhysCanvas organizes multi-view RGB, tactile images, marker displacement, and task state into a unified multimodal canvas. All modalities share a video VAE and pretrained generative backbone, allowing optical touch to reuse visual and temporal priors. Motion images derived through forward kinematics and camera projection provide spatially aligned action conditioning, while an initial RGB anchor and updated reference observations guide sequential generation. Our consistency-aware post-training uses the frozen Scene Reward Model, Modality Reward Model, and State Reward Model to supervise scene preservation, visual–tactile compatibility, and agreement with real task outcomes. The resulting world model supports closed-loop policy evaluation and provides imagined interactions and rewards for GRPO policy updates. Across four real-world manipulation tasks, State Reward Model supervision increases the arithmetic mean of task-wise Pearson correlations between model and real-world success rates from 0.602 to 0.885 under fixed-action replay. Policy post-training in the reward-guided world model further increases π0.5’s mean real-world success rate from a 59.1% SFT baseline to 70.8%. These results demonstrate the value of unified multimodal modeling and consistency-aware post-training for robot policy evaluation and learning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, World Model, Vision Tactile Fusion

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HgsiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PhysCanvas is an action-conditioned visual–tactile world model for robot policy evaluation and learning. It represents multi-view RGB, tactile images, marker displacement, and task state in a unified canvas processed by a shared pretrained video backbone. Consistency-aware post-training uses scene, modality, and state reward models to encourage scene preservation, visual–tactile compatibility, and agreement with observed task outcomes. Across four manipulation tasks, adding state supervision to the model with scene-consistency supervision (M2 → M3) increases the reported mean task-wise correlation between simulated and real-world success rates from 0.602 to 0.885. Policy optimization in the full simulator improves π0.5’s mean real-world success rate from approximately 59.1% after SFT to 70.8%, a gain of 11.7 percentage points. The additional gain over policy optimization in the world model without consistency post-training is approximately 3.3 percentage points, from 67.5% to 70.8%.

Critical strengths
  1. An end-to-end pipeline with real-world validation. The paper connects multimodal world-model training, closed-loop imagined interaction, and policy optimization to evaluation on a physical robot. The reported improvements over SFT provide evidence that learning in the simulator can yield benefits beyond its internal reward scores.
  2. A unified representation that reuses video pretraining. The canvas places visual observations, tactile images, marker displacement, and task state within a shared visual representation. This provides a concrete way to reuse pretrained video models while generating the tactile displacement signals consumed by the evaluated policies.
  3. Consistency supervision with measurable benefits. The state objective is designed to match both successful and unsuccessful outcomes rather than uniformly favoring successful outcomes. Separately, adding modality-consistency supervision (M3 → M4) reduces all three marker-displacement errors, although contact-event F1 does not improve. The full consistency-post-trained simulator also improves average downstream performance, with gains varying across tasks.
Critical weaknesses
  1. The policy-evaluation protocol does not fully establish reliability. Figure 6 does not clearly specify how simulated success is determined. If it uses the State RM employed during training, independent validation is needed to distinguish improved outcome prediction from adaptation to that evaluator. Moreover, correlations pooled across three policy families do not establish reliable comparison of different versions of the same policy.
  2. The contributions of touch and the unified canvas are not isolated. The paper does not report RGB-only or separate tactile-branch comparisons under matched conditions, particularly for policy evaluation and real-world policy improvement. The experiments demonstrate the feasibility of the complete multimodal system, but do not establish the incremental benefit of tactile inputs or the advantage of the shared canvas over alternative architectures.
  3. The stability of the additional real-world gain is difficult to assess. Table 4 reports means and population standard deviations over three runs, but does not clarify whether these represent independent training runs or repeated evaluations of a fixed policy. The number of real-world trials for this experiment is also not explicitly stated. Consequently, the approximately 3.3-percentage-point average gain over the ordinary world-model simulator is difficult to assess relative to training and evaluation variability.
Questions for authors
  1. How is success determined for the simulated rollouts in Figure 6? If the State RM used during training also provides these judgments, is there any independent validation, such as human assessment, of the simulated outcomes?
  2. Do you have any existing RGB-only or separate tactile-branch comparison results, particularly for policy evaluation or real-world policy improvement, that could help isolate the contributions of tactile inputs and the unified canvas?
  3. What do the three runs in Table 4 represent: independent policy-training runs or repeated evaluations of a fixed trained policy? Please also clarify the number of real-world trials per task and method in each run, whether methods were evaluated under the same initial conditions, and how the evaluated checkpoints were selected.
5 comments

W1 & Q1 Policy evaluation protocol

Evaluation independent of the reward models. As stated in lines 427–428 of the manuscript, we do not use the State Reward Model to determine evaluation outcomes. Instead, three human experts assess the evaluation samples, which are randomly assigned among them to reduce systematic assignment bias. Thus, the evaluation does not directly reuse the reward-model predictions employed during post-training.

Policy-specific correlations. Table H1 reports evaluation correlations separately for each policy. Pi0.5 has a lower correlation than ACT and FM. As shown in Table H2, Pi0.5 generally exhibits larger end-effector motion spans along the X and Y axes, suggesting a broader range of motion. Its actions may therefore be less well covered by the world model’s training distribution, potentially increasing rollout errors. In future versions, we will expand the action coverage of the training data to better accommodate diverse policy behaviors.

Table H1. Policy-specific evaluation correlations.

MetricACTFMPi0.5
Correlation coefficient0.81750.87210.4685

Table H2. Mean end-effector motion spans across ACT, FM and Pi0.5.

Model TypeRobot ArmX SpanY SpanZ Span
ACTLeft0.078770.224730.24524
ACTRight0.126630.247490.16259
FMLeft0.074420.215800.23988
FMRight0.124910.239570.16130
Pi0.5Left0.088860.235040.24404
Pi0.5Right0.168480.266550.15848

W2 & Q2 Ablation of the tactile branch

Thank you for the suggestion. We examine tactile input from three perspectives: visual generation quality (Table H3), world-model-based policy evaluation (Table H4), and policy improvement through RL (Table H5). Because the latter two require substantial real-robot testing and human assessment, we focus on the Test Tube and Bottle tasks.

Visual generation quality. Removing tactile input reduces performance on several visual metrics, including PSNR, SSIM, LPIPS, FID, and FVD. Nevertheless, RGB-only PhysCanvas remains competitive and outperforms all listed baselines on these five metrics (Table H3).

World-model-based policy evaluation. In Table H4, we average the three batches within each task–policy setting and compute the Pearson correlation between world-model and real-world success rates across the six resulting pairs. The correlation is 0.76 for RGB-only and 0.90 for RGB+Tactile, indicating closer agreement with real-world success rates when tactile input is included.

Policy improvement through RL. Table H5 reports real-world success rates for SFT and RL policies with and without tactile input. Averaged across the four task–policy settings, RGB-only RL improves success from 37.48% to 42.48%, a gain of 5.00 percentage points. RGB+Tactile RL improves success from 39.58% to 45.40%, a gain of 5.83 percentage points. RGB+Tactile RL achieves a higher mean success rate than RGB-only RL (45.40% vs. 42.48%), a difference of 2.93 percentage points. These results suggest that tactile input can improve downstream policy performance, while RL remains beneficial without tactile input.

Table H3. Visual generation metrics for RGB-only PhysCanvas, RGB+Tactile PhysCanvas, and baselines.

ModelPSNR ↑SSIM ↑LPIPS ↓FID ↓FVD ↓Subject ↑Bg. ↑Smooth ↑EEF (px) ↓Object (px) ↓
EVAC16.4390.5280.25826.512165.70579.22390.45697.7311.9642.15
Open-Sora17.9160.5990.22834.762183.8678.51990.23498.1571.4251.426
WoVR17.2780.59720.20326.086138.24380.24791.08298.0592.7921.164
PhysCanvas RGB-only20.5020.710.14124.812128.10380.24890.41998.390.8441.438
PhysCanvas RGB+Tactile21.1720.7350.12121.459116.27480.34990.56598.3010.6851.15

Table H4. Tactile-input ablation for world-model-based policy evaluation.

MethodTest Tube, ACTTest Tube, FMTest Tube, Pi0.5Bottle, ACTBottle, FMBottle, Pi0.5
Real World (RGB only)15%, 5%, 0%0%, 0%, 0%45%, 55%, 45%50%, 40%, 40%20%, 15%, 25%80%, 80%, 85%
Eval in WM (RGB only)40%, 10%, 10%5%, 30%, 20%55%, 70%, 70%30%, 25%, 30%45%, 30%, 45%55%, 50%, 60%
Real World (RGB+Tactile)15%, 10%, 0%0%, 0%, 0%45%, 60%, 45%55%, 30%, 35%20%, 15%, 25%85%, 85%, 95%
Eval in WM (RGB+Tactile)30%, 15%, 20%5%, 25%, 15%60%, 60%, 55%50%, 20%, 25%35%, 25%, 40%50%, 75%, 55%

Table H5. Real-world success rates (%) of SFT and RL policies with and without tactile input.

MethodTest Tube, FMTest Tube, Pi0.5Bottle, FMBottle, Pi0.5
RGB-only SFT0.0±0.00.0 \pm 0.048.3±4.748.3 \pm 4.720.0±4.020.0 \pm 4.081.6±2.381.6 \pm 2.3
RGB-only RL0.0±0.00.0 \pm 0.055.0±4.055.0 \pm 4.028.3±8.528.3 \pm 8.586.6±2.386.6 \pm 2.3
RGB+Tactile SFT0.0±0.00.0 \pm 0.050.0±7.050.0 \pm 7.020.0±4.020.0 \pm 4.088.3±4.788.3 \pm 4.7
RGB+Tactile RL0.0±0.00.0 \pm 0.060.0±8.160.0 \pm 8.135.0±4.035.0 \pm 4.086.6±2.386.6 \pm 2.3

W3 & Q3 Experimental protocol and policy improvement

Real-world evaluation protocol and checkpoint sensitivity. We apologize for not describing the experimental settings clearly. For each task and policy, we evaluate a fixed checkpoint over 60 real-world trials, divided into three batches of 20 trials conducted during separate, fixed time periods. We use default RL checkpoints obtained after 20 RL updates. Table H6 additionally examines checkpoints obtained after 20, 30, and 40 updates on the Test Tube and Bottle tasks. The average success rate increases slightly, which we attribute to normal variation across checkpoints.

Table H6. Real-world policy success rates with default and multiple RL checkpoints.

MethodTest Tube, FMTest Tube, Pi0.5Bottle, FMBottle, Pi0.5
RL in WM (default)0.0±0.00.0 \pm 0.060.0±4.060.0 \pm 4.033.3±2.333.3 \pm 2.390.0±0.090.0 \pm 0.0
RL in WM+RM (default)0.0±0.00.0 \pm 0.060.0±8.160.0 \pm 8.135.0±4.035.0 \pm 4.086.6±2.386.6 \pm 2.3
RL in WM (multi ckpts)0.0±0.00.0 \pm 0.060.0±4.060.0 \pm 4.035.0±8.135.0 \pm 8.190.0±4.090.0 \pm 4.0
RL in WM+RM (multi ckpts)0.0±0.00.0 \pm 0.061.6±6.261.6 \pm 6.238.3±10.238.3 \pm 10.290.0±8.190.0 \pm 8.1

Modest additional gains for Pi0.5. In the aggregate results cited from the manuscript, Pi0.5 improves from 59.17% to 67.50% after RL, a gain of 8.33 percentage points. The additional reported gain from “RL in WM” to “RL in WM+RM” is 3.3 percentage points. Pi0.5 already achieves a relatively high success rate after RL in the standard world model. Further improvements in world-model generation quality may therefore yield diminishing returns, resulting in only modest gains in policy success rate.

Summary

Thank you again for your constructive feedback. Our responses address three main points:

  • Independent evaluation: We clarify human assessment independent of reward-model predictions and report policy-specific correlations (Table H1).
  • Tactile contribution: Tactile input improves several generation metrics, evaluation correlation, and mean RL success (Tables H3–H5).
  • Protocol and robustness: We clarify the 60-trial protocol, examine checkpoints at 20/30/40 RL updates, and discuss checkpoint variation and diminishing returns for Pi0.5 (Table H6).

We hope these clarifications and additional results address your concerns, and we welcome further discussion.

Reviewer_bKt1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PhysCanvas, a world model that jointly predicts visual and tactile observations for robot manipulation. It uses consistency-based training to improve predictions and supports policy evaluation and learning through simulated interactions. Experiments on four real-world tasks show improved agreement with real-world outcomes and higher robot policy success rates.

Critical strengths
  1. A coherent framework for multimodal robot simulation. The unified canvas allows visual observations, tactile images, marker displacements, and task states to share a pretrained video backbone. Together with explicit consistency supervision, this offers a practical way to adapt video models to contact-rich manipulation and provides a useful system contribution.
  2. Validation through downstream evaluation and real-world policy improvement. The paper evaluates usefulness beyond visual prediction quality. State supervision improves the mean task-wise correlation between simulated and real-world success rates from 0.602 to 0.885, while policy post-training increases π₀.₅’s average real-world success rate from 59.1% to 70.8% across four tasks. These results provide substantive evidence that the framework can support robot policy development.
Critical weaknesses
  1. Some key experimental comparisons are missing. Figure 6 compares M2 and M3, but does not report results for the final M4 model, leaving its policy-evaluation performance unclear. Also, there is no controlled comparison that isolates the benefit of the unified canvas over other ways of combining vision and touch.
  2. It is still unclear to me about the reliability of the reward models. The paper describes how the Scene, Modality, and State Reward Models are trained, but does not report their standalone performance. Since these models guide world-model post-training, and the State Reward Model also supplies rewards for policy learning, their errors could affect both stages. Reporting performance on held-out, human-labeled real and generated samples would help establish their reliability.
Questions for authors

My questions mainly follow the weaknesses above, especially whether the unified canvas design itself provides a clear benefit.

3 comments

We thank the reviewer for the detailed and constructive feedback. The comments have helped us improve the presentation of the ablations and the evaluation of Reward Model reliability. We respond to each point below and will incorporate these clarifications into the manuscript.

W1 Additional ablation studies

Ablation of M4M_{4} for policy evaluation. Thank you for pointing this out. We further compare M3M_{3} and M4M_{4} in Table B1. Each cell reports success rates for three chronological batches of 20 trials each. Within each task–policy setting, batches are aligned in the same order across rows. M4M_{4} further improves generation quality through modality alignment, while providing a relatively modest improvement in evaluation correlation over M3M_{3} (from 0.885 to 0.917). This suggests that the State Reward Model accounts for most of the improvement in task-level correlation (from 0.602 for M2M_{2} to 0.885 for M3M_{3}).

Table B1. Real-world and world-model success rates for M3M_{3} and M4M_{4}. Each cell contains three temporally ordered batches of 20 trials.

ModelTest Tube, ACTTest Tube, FMTest Tube, Pi0.5Bottle, ACTBottle, FMBottle, Pi0.5Blocks, ACTBlocks, FMBlocks, Pi0.5Rubik’s Cube, ACTRubik’s Cube, FMRubik’s Cube, Pi0.5
Real World15%, 10%, 0%0%, 0%, 0%45%, 60%, 45%55%, 30%, 35%20%, 15%, 25%85%, 85%, 95%40%, 35%, 40%40%, 20%, 15%45%, 50%, 45%0%, 10%, 5%0%, 0%, 0%60%, 45%, 50%
PhysCanvas M3M_{3}30%, 15%, 20%5%, 25%, 15%60%, 60%, 55%50%, 20%, 25%35%, 25%, 40%50%, 75%, 55%30%, 20%, 30%45%, 25%, 20%45%, 55%, 45%0%, 10%, 5%0%, 0%, 0%60%, 35%, 40%
PhysCanvas M4M_{4}15%, 15%, 5%5%, 10%, 15%70%, 60%, 55%50%, 35%, 25%30%, 15%, 40%50%, 90%, 70%45%, 20%, 35%30%, 10%, 20%45%, 55%, 45%0%, 10%, 0%5%, 0%, 0%70%, 50%, 50%

Ablation of multimodal fusion. We compare our canvas-based representation with a fusion baseline inspired by VT-WM [Higuera et al., 2026]. VT-WM uses separate visual and tactile encoders and combines their tokens through spatial and temporal attention in a transformer predictor. We adapt this fusion strategy for comparison with PhysCanvas M1M_{1}, before reward-model post-training. As shown in Tables B2 and B3, PhysCanvas M1M_{1} outperforms the VT-WM-inspired baseline on most evaluated metrics.

Table B2. Visual metrics for PhysCanvas M1M_{1} and the VT-WM-inspired fusion baseline.

ModelPSNR ↑SSIM ↑LPIPS ↓FID ↓FVD ↓Subject ↑Bg. ↑Smooth ↑EEF (px) ↓Object (px) ↓
VT-WM-inspired fusion18.1210.6310.15434.602163.79180.45390.25498.1321.4731.973
PhysCanvas M1M_{1}19.8690.6600.11333.145154.90180.27290.34798.3330.8101.651

Table B3. Tactile metrics for PhysCanvas M1M_{1} and the VT-WM-inspired fusion baseline.

ModelPSNR ↑SSIM ↑LPIPS ↓Contact MAE ↓Onset ↓Offset ↓F1 ↑Active MAE ↓Idle FP-MAE ↓Amp.MAE ↓
VT-WM-inspired fusion35.3710.9250.1460.0281.6831.0470.3645.2141.2248.281
PhysCanvas M1M_{1}36.2920.9310.1270.0261.1820.9870.3494.8340.9007.634

W2 Reliability of the reward models

We appreciate the reviewer’s careful assessment and suggestions. Table B4 evaluates the reliability of the State, Scene, and Modality Reward Models on real and world-model-generated data. The Scene Reward Model has lower classification accuracy than the other two models. For scene consistency, we treat hallucination-induced changes in the position or shape of key objects as inconsistencies. We hypothesize that this criterion involves more ambiguous cases than task-completion or modality-consistency assessment. Accuracy decreases on world-model-generated data by 2.23, 3.08, and 1.59 percentage points for the State, Scene, and Modality Reward Models, respectively.

Table B4. Reliability of the three reward models on real data (top) and world-model-generated data (bottom).

Model & ConfigurationAccuracyPrecisionRecallF1AUROC
State RM — Real Data98.73%94.85%99.10%96.90%99.90%
Scene RM — Real Data91.00%98.00%87.00%92.20%99.00%
Modality RM — Real Data94.59%91.25%93.58%92.40%99.13%
Model & ConfigurationAccuracyPrecisionRecallF1AUROC
State RM — WM-Generated Data96.50%94.00%97.50%95.70%99.50%
Scene RM — WM-Generated Data87.92%99.76%81.75%89.86%99.07%
Modality RM — WM-Generated Data93.00%91.50%95.00%93.20%99.00%

Summary

Thank you again for your constructive feedback. Our responses address three main points:

  • Modality alignment: M4M_{4} improves evaluation correlation over M3 from 0.885 to 0.917 (Table B1).
  • Fusion design: PhysCanvas M1M_{1} outperforms the VT-WM-inspired baseline on most evaluated metrics (Tables B2 and B3).
  • Reward-model reliability: We evaluate all three reward models on real and generated data and discuss their accuracy differences (Table B4).

We hope these clarifications and additional results address your concerns, and we welcome further discussion.

References

Carolina Higuera, Sergio Arnaud, Byron Boots, Mustafa Mukadam, Francois Robert Hogan, and Franziska Meier. “Visuo-Tactile World Models.” arXiv:2602.06001 2026.

Dear Reviewer bKt1,

Thank you once again for your thoughtful review. We have added policy-evaluation results for the full M4 model, compared the unified canvas with a separate-encoder fusion baseline, and evaluated all three reward models on both real and generated observations.

We have submitted the revised manuscript, incorporating the corresponding clarifications and additional experimental results to address the issues raised in your review. All revisions are highlighted in blue for ease of reference.

With the discussion period ending on November 18, 2026 (AoE), please feel free to let us know if there is anything else you would like us to clarify before then.

We sincerely appreciate your time and feedback.

Best regards,
The authors of Submission 3366

Reviewer_y5RCOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

Authors introduce a visuo-tactile world model for contact-rich manipulation by using a unified video backbone of the RGB image, tactile image, and so on. This model is used for policy evaluation and policy post-training. While the idea is interesting and the result is impressive, the presentation and experiment design can be clearly improved.

Critical strengths
  • Impressive prediction quality. The prediction video shown in the supp is quite impressive. In addition, being able to predict tactile information accurately is also impressive
  • Comprehensive experiment results including policy evaluation and post-training in the real world
  • A lot of detailed information in the appendix for reproducing
Critical weaknesses
  • The definition of state-consistent is not super clear. Authors highlight state-consistent visuo-tactile world model in the abstract and introduction. However, it is not super clear what do authors mean by state-consistent. Typically robotics community refers to state as the privileged information (e.g. object pose, and so on). Does authors refer to unified latent space as consistent state?
  • Inconsistency annotation. In section 3.1, the observation is annotated as o (e.g. o^cam_t), while the observation is annotated as I in section 3.2 (e.g. I^cam_t). Unless they actually refer to different things. It should be unified.
  • Writing can be further improved. My confusion is listed in the question section.
  • Evaluation results seem to be quite noisy, including test tube and bottle. There seems to be a lot of false positive in the world model.
  • The comparisons with the baseline are not entirely fair. The paper uses tactile sensor. However, the baseline does not choose the world model with tactile sensor, which is not an entire apple-to-apple comparison. This might be one of the biggest concern I have
Questions for authors
  • What is g_state exactly in Section 3.2? It was not defined or explained.
  • I am not super clear about the definition of the task state. In Figure 10 caption, it sounds like a 2-dimension vector like (True, False), as indicated by "Each row shows state 0 (the success condition is not satisfied) and state 1 (the success condition is satisfied) for one task." But it seems that the RGB derived from such state, i.e. g_state(s_t), is very arbitrary. For example, why the state 0 image is RGB (18, 72, 45) and state 1 image is RGB (91, 224, 104) in Test Tube? Why not RGB (0, 0, 0) and RGB (255, 255, 255). Also, the image does not look like a pure solid color. It appears to have some color patterns, but why the caption seems to indicate that this is a solid color.
  • In section 3.2 Action-conditioned latent dynamics, what is S_state exactly? Seems to be different?
  • Equation (7) is a bit weird. I assume that v_theta is predicted velocity for denoising process. Does Equation (7) mean that this is a one-step diffusion model, as \tau is just one value, instead of iterative denoising like z_t -> z_{t-1} -> z_{t-2} -> z_0
  • During GRPO policy training, is it possible that the policy will hack the reward function? I think RL post training in world model is still a problem under exploration. It is interesting to see that authors make it work without too much tuning.
  • Section 4.3 first paragraph piles all the number, which might not be the best way of presentation. It will be more informative to include more analysis in the experiment section.
  • What is Onset / Offset in Table 3? While it is totally to skip the definition of the commonly used evaluation metrics like PSNR and so on. Some less well-known metrics like onset and offset might need some explanation
  • For Table 4, what is RL in WM+RM? Is RM reward model? If yes, the abbreviation is not mentioned in the paper, such as Reward Model (RM). In addition, what is the difference between RL in WM and RL in WM+RM? I might miss the explanation but I tried to find the explanation but failed.
4 comments

We sincerely appreciate the reviewer’s careful reading and constructive feedback. The comments have helped us clarify the task-state representation, improve the presentation, and explain the evaluation more clearly. We address each point below and will incorporate the corresponding clarifications into the manuscript.

W1 & Q2 Definitions of state consistency and task state

Definition of state consistency. We apologize for the unclear terminology. Here, “state” refers to the task’s completion status. State consistency means that the predicted task state agrees with the corresponding real-world task state. We introduce auxiliary task-state supervision to encourage the world model to preserve task-relevant semantic information in addition to visual similarity.

Definition of task state. We represent the task state as a 32 × 672 textured color block that encodes both task identity and completion status, as illustrated in Appendix Figure 10. This block is concatenated with the other modalities on the canvas, as shown in Figure 2 of the main text.

The task-state region is therefore not a solid-color block. The RGB triplets in the caption of Figure 10 specify its dominant colors. Vertical stripes indicate an incomplete task, whereas horizontal stripes indicate a completed task. Colors and textures distinguish task identities and completion states.

Choice of task-state colors. Our goal is to make different task-state categories distinguishable, making it easier for the network to distinguish between them. This design is similar to Hyperspherical Prototype Networks [Mettes et al., 2019], which explicitly separates class prototypes in the output space. In our canvas, we pursue this goal through distinct color families for different tasks and different shades and textures for completion states.

To examine this design, we conduct a controlled experiment that requires predicting task state from visual observations. Specifically, we train a simple MLP to map the latent representations of three camera views to the task-state representation, using 1,000 training examples and 150 test examples. We compare six encoding schemes in Table Y1. Distinct colors with texture achieve 72.00% accuracy, compared with 70.00% for similar colors with texture and 69.33% for distinct colors alone. The black + white alternative achieves 70.67%, while black + distinct colors + texture achieves the highest accuracy of 74.67%. These results suggest that distinguishable colors and textures can facilitate task-state prediction in this controlled setting, without implying that our specific RGB values are optimal.

Table Y1. Ablation of task-state representations.

MetricsDistinct colorsDistinct colors + texture (default)Similar colors + textureBlack + whiteBlack + distinct colorsBlack + distinct colors + texture
Accuracy69.33%72.00%70.00%70.67%73.33%74.67%

W2 Inconsistent notation

Thank you for pointing this out. We used inconsistent notation across the two sections. We will follow Section 3.2 and consistently denote the visual observations by ItcamI^{\mathrm{cam}}_t.

W3 Writing and presentation

We sincerely thank the reviewer for the careful examination of the manuscript. We will address each of the identified issues.

W4 Success-rate discrepancies in evaluation

Table Y2 shows discrepancies between real-world and world-model success rates for ACT and FM on the Test Tube and Bottle tasks. The M2 world model overestimates success in three of the four settings and underestimates it for ACT on the Bottle task. We hypothesize that the small size of the relevant objects, together with the transparency of the Test Tube, makes their interactions difficult to model. After post-training with the State Reward Model, the mean absolute success-rate gap across these four settings decreases from 32.50 to 12.09 percentage points. The mean signed gap (world model minus real world) decreases from 22.50 to 7.92 percentage points. These results indicate better agreement with real-world success rates.

Table Y2. Policy success rates in the real world and in world models before and after State Reward Model post-training.

MethodTest Tube (Real)Test Tube (WM)Test Tube (WM+RM)Bottle (Real)Bottle (WM)Bottle (WM+RM)
ACT8.33%60.00%21.67%40.00%20.00%31.66%
FM0.00%23.33%15.00%20.00%55.00%31.67%

W5 Fair comparison with baselines

Table Y3 reports a tactile-input ablation. Removing tactile input slightly degrades the visual generation metrics overall. However, even without tactile input, PhysCanvas outperforms all three listed baselines on most metrics. At the same time, its higher object tracking error suggests that removing tactile feedback may weaken the model’s ability to capture object motion during contact-rich interactions.

Table Y3. Visual generation metrics for RGB-only PhysCanvas, RGB+Tactile PhysCanvas, and baselines.

ModelPSNR ↑SSIM ↑LPIPS ↓FID ↓FVD ↓Subject ↑Bg. ↑Smooth ↑EEF (px) ↓Object (px) ↓
EVAC16.4390.5280.25826.512165.70579.22390.45697.7311.9642.15
Open-Sora17.9160.5990.22834.762183.8678.51990.23498.1571.4251.426
WoVR17.2780.59720.20326.086138.24380.24791.08298.0592.7921.164
PhysCanvas RGB-only20.5020.710.14124.812128.10380.24890.41998.390.8441.438
PhysCanvas RGB+Tactile21.1720.7350.12121.459116.27480.34990.56598.3010.6851.15

Q1 Definition of gstateg_{\mathrm{state}}

gstateg_{state} is a rule-based mapper that maps task type and completion status to a colored texture representation.

Q3 Definition of SstateS_{\mathrm{state}}

SstateS_{state} denotes the set of positional indices identifying the task-state tokens in the latent representation of the full canvas.

Q4 Clarification of Equation 7

During post-training, we use a single denoising step, with the diffusion time τ\tau randomly sampled from [0, 1]. This choice balances training efficiency and the cost of denoising, since unrolling multiple denoising steps would substantially increase the computational cost of post-training.

Q5 Reward hacking during GRPO

Thank you for the positive assessment. We agree that reward hacking is a potential concern during GRPO training Gao et al., 2023. We take two measures to mitigate this risk. First, we seek to provide accurate supervision through task-specific State Reward Models which provide reward for GRPO and report their validation accuracy in Table Y4. Second, we limit the duration of policy post-training, using at most 40 RL updates (similar with WovR[Jiang et al., 2026]) to reduce the risk of overoptimization.

Table Y4. Validation performance of task-specific State Reward Models.

ModelAccuracyPrecisionRecallF1AUROC
Test Tube98.73%94.25%98.92%96.53%99.92%
Bottle98.14%94.43%98.44%96.39%99.89%
Blocks99.67%99.22%99.33%99.28%99.96%
Rubik’s Cube98.36%91.48%99.70%95.42%99.90%

Q6 Definition of onset and offset

We define these metrics in the “Event F1 and boundary timing” section of Appendix F. Onset is the time at which tactile contact begins, and offset is the time at which it ends. The onset and offset errors measure the timing discrepancies between predicted and ground-truth contact transitions.

Q7 Definition of "WM+RM"

We apologize for the unclear description. RM stands for reward model. “RL in WM+RM” denotes policy reinforcement learning in a world model that has been post-trained using the three reward models.

Summary

Thank you again for your constructive feedback. Our responses address three main points:

  • Definitions and writing: We clarify state consistency, task-state encoding, notation, and methodological details, and add a color choice ablation (Table Y1).
  • Fair baseline comparison: RGB-only PhysCanvas outperforms all listed baselines on PSNR, SSIM, LPIPS, FID, and FVD without tactile input (Table Y3).
  • False positives in evaluation: We discuss success-rate overestimation and show that State Reward Model post-training reduces the mean absolute success-rate gap from 32.50 to 12.09 percentage points (Table Y2).
  • Reward Hacking during GRPO: We discussed how we mitigate this issue by ensuring the accuracy of the reward model and controlling the number of RL updates. (Table Y4).

We hope these clarifications and additional results address your concerns, and we welcome further discussion.

References

Leo Gao, John Schulman, and Jacob Hilton. Scaling Laws for Reward Model Overoptimization.. ICML 2023.

Mettes, P., van der Pol, E., and Snoek, C. G. M. Hyperspherical Prototype Networks. NeurIPS 2019.

Jiang, Zhennan, Shangqing Zhou, Yutong Jiang, Zefang Huang, Mingjie Wei, et al. WoVR: World Models as Reliable Simulators for Post-Training VLA Policies with RL. CoRL 2026.

Dear Reviewer y5RC,

Thank you once again for your thoughtful review. We have clarified the task-state definitions and methodological details, added a fair RGB-only baseline comparison, analyzed false positives, and explained our measures to mitigate reward hacking during GRPO.

We have also submitted the revised manuscript, incorporating the corresponding clarifications and additional experimental results to address the issues raised in your review. All revisions are highlighted in blue for ease of reference.

With the discussion period ending on November 18, 2026 (AoE), please feel free to let us know if there is anything else you would like us to clarify before then.

We sincerely appreciate your time and feedback.

Best regards,
The authors of Submission 3366

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.