acceptodds
Under review as a conference paper at ICLR 2027

DreamChaser: A Self-Improving Framework for World Action Models via Video-guided Action Refinement

Abstract

World Action Models (WAMs) built on pre-trained video generators learn video prediction and action generation under asymmetric supervision. The video branch inherits dense spatiotemporal supervision from large-scale video pre-training, whereas the action branch relies on comparatively limited paired action annotations. We hypothesize that this asymmetry in data scale contributes to a capability gap: a WAM can predict plausible task outcomes that its generated actions fail to realize. We introduce DreamChaser, a self-improving framework that uses the model’s video predictions to guide action refinement and acquire additional action supervision through interaction. A lightweight refiner, trained solely on the original demonstrations, uses predicted visual outcomes as targets for correcting local execution deviations. DreamChaser selectively aggregates the resulting rollout data and combines them with the original demonstrations to fine-tune the WAM. This process converts video-guided corrections into additional action supervision, forming an iterative loop of interaction and policy updates without further human demonstrations or an external expert during rollout. On LIBERO, DreamChaser fine-tuning increases FastWAM-joint’s average success rate from 97.3% to 97.9% when evaluated without the refiner. On LIBERO-plus, three rounds of self-improvement increase the mean task success from 72.1% to 80.1% across ten tasks. On a real AgileX PiperX, DreamChaser improves the FastWAM-joint average success score on cup sleeving, plate placement, and bottle straightening from 27.0% to 41.7% after two online rounds.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Models, Embodied AI, Robotics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_b9uTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a self-improving framework for world action models by utilizing an action refiner to correct local execution deviations. As world action models usually exhibit a stronger video prior than their action generator, this framework triggers action refinement when execution deviates from the predicted visual goal, and combines the corrective trajectories with expert data for further improvement of the WAM. Experiments show the refiner enables performance gains while providing corrective data to support further model finetuning. Overall, the framework demonstrates effective self-improvement over both simulated and real-robot settings.

Critical strengths
  1. The proposed framework utilizes the stronger prior from video prediction to guide the local correction with a lightweight action refiner, which in general can be integrated with other existing world action models.
  2. The experiments provide extensive evaluations in the simulated and real-robot environments and showcase a consistent performance trend.
Critical weaknesses
  1. In Table 1, it seems that all compared methods are fairly saturated on the LIBERO tasks. The FastWAM-joint baseline is merely 0.7 points below the best-performing method on average. The comparisons are less informative without understanding the detailed evaluation protocol (e.g., number of trials and seeds) and the standard error/deviation of the reported numbers.
  2. The simulated evaluation lacks comparisons with external baselines, such as Latent Policy Barrier[1] or CCIL[2], which can be compared against the proposed WAM+action refiner. It also remains unclear how the proposed self-improvement framework compares against existing self-improving methods (e.g., in terms of efficiency).
  3. Further clarifications are required for the questions below, particularly for the first two questions.

[1] Sun et al. Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution. 2025

[2] Ke et al. CCIL: Continuity-based data augmentation for corrective imitation learning. 2023

Questions for authors
  1. What target-state control interface did the authors use in the action refiner? Is it an IK solver? How does the refiner avoid obstacle collisions during corrective execution?
  2. Given the action refiner formulation, would it be able to serve beyond simply "refinement" purposes? How does the performance compare to the proposed method, if the action chunk predicted by WAM is discarded, and the refiner is used to directly extract the target robot state from the predicted visual goal, which is then fed into the control interface for execution?
  3. How are HRH_R and HH decided for the experiments? How does the relation between HRH_R and HH affect the performance?
  4. Is the action refiner (i.e., the target predictor) frozen during finetuning stages? Would finetuning action refiner bring extra benefits?
  5. In Figure 3, typo “color gitter” -> “color jitter”
7 comments

Dear Reviewer b9uT,

Thank you for your thorough review and constructive feedback. We have carefully considered your suggestions and respond in detail below.

W1. LIBERO benchmark saturation, evaluation protocol, and variability of reported results

Thank you for highlighting benchmark saturation and evaluation variability. We evaluated LIBERO over five runs with different random seeds, using 50 trials per task in each run, and report mean success, sample standard deviation, and gains over FastWAM (Table R1).

• Table R1. Repeated LIBERO evaluation
MethodSR: mean ± SD (%)Gain vs FastWAM (pp)
FastWAM97.32 ± 0.080.00
FastWAM + DreamChaser refiner97.80 ± 0.14+0.48
FastWAM + Vanilla-FT96.46 ± 0.11-0.86
FastWAM + DreamChaser97.90 ± 0.14+0.58
FastWAM + DreamChaser + refiner98.04 ± 0.15+0.72

Most LIBERO tasks already approach 100% success. Refinement improves performance on tasks that are not yet saturated. The table below highlights the two tasks with the lowest baseline success as examples, with mean success increasing from 89% to 94% (Table R2).

• Table R2. Examples of refinement gains on LIBERO
Task (SR, %)FastWAMFastWAM + DreamChaser refiner
Wine bottle to cabinet top88.094.0
Black bowl: top drawer to plate90.094.0
Mean89.094.0

To assess broader applicability, we evaluate all 24 tasks of the more challenging RoboCasa-GR1 benchmark. Three improvement rounds increase success from 67.8% to 72.0%, compared with 62.2% for vanilla fine-tuning (Table R3). These results support DreamChaser's broad applicability across diverse manipulation tasks and benchmarks.

• Table R3. Full RoboCasa-GR1 evaluation
Method (SR, %)InitialInitial + refinerR1R1 + refinerR2R2 + refinerR3R3 + refiner
FastWAM + Vanilla-FT67.8—65.1—63.3—62.2—
FastWAM + DreamChaser67.868.569.670.271.071.572.072.4
Note: R1–R3: self-improvement rounds.

W2. Comparisons with LPB and CCIL for policy robustness and with self-improvement methods for effectiveness and efficiency

Thank you for these valuable suggestions. LPB[1] and CCIL[2] provide complementary approaches to policy robustness through latent-space guidance and corrective data augmentation. We compare both methods on adapted LIBERO-Plus and the three real-robot tasks with the same FastWAM backbone (Table R4). DreamChaser's refiner achieves stronger task performance with less additional training than either baseline, while adding substantially less inference overhead than LPB. These results highlight the effectiveness and efficiency of visual-goal-guided local correction within WAMs.

• Table R4. External policy-robustness baselines
MethodLP InitialLP SR (%)LP Added (ms)LP Extra train (h)Real InitialReal score (0–100)Real Added (ms)Real Extra train (h)
FastWAM + LPB72.174.552.41.427.033.562.70.5
FastWAM + CCIL72.174.20.05.427.037.80.03.8
FastWAM + DreamChaser refiner72.176.811.30.927.043.715.40.2
Note: LP: adapted LIBERO-Plus SR (%). Real: normalized score (0–100). Added: latency per correction. Extra train: additional training time.

Compared with self-improvement methods SILVR[3] and Q-Planning[4], DreamChaser achieves larger policy gains with shorter data-collection and filtering time (Table R5). The updated WAM retains these gains when deployed without the refiner. These results demonstrate more effective and efficient self-improvement across both settings. We will include further experimental comparisons and discussion with LPB[1], CCIL[2], SILVR[3], and Q-Planning[4] in the revised manuscript to better position DreamChaser among robust control and robot self-improvement methods.

• Table R5. Comparison with self-improvement methods
MethodLP InitialLP R1LP R2LP R3LP Collect (h)Real InitialReal R1Real R2Real Collect (h)
FastWAM + Vanilla-FT72.171.470.970.21.927.023.719.23.2
FastWAM + SILVR72.172.372.873.110.227.028.630.614.2
FastWAM + Q-Planning72.174.275.176.59.827.032.335.311.4
FastWAM + DreamChaser72.176.877.880.11.927.036.341.73.2
Note: Collect time: mean collection and filtering time per round. Training time is matched across methods for fair comparison.

References

[1] Zhanyi Sun and Shuran Song. Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution. NeurIPS, 2025.

[2] Liyiming Ke et al. CCIL: Continuity-Based Data Augmentation for Corrective Imitation Learning. ICLR, 2024.

[3] Calvin Luo et al. Self-Improving Loops for Visual Robotic Planning. ICLR, 2026.

[4] Varun Giridhar et al. Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning. arXiv:2608.21204, 2026.

Q1. How are the refiner's target states executed, and how are obstacle collisions avoided?

Thank you for this valuable question. A key advantage of DreamChaser is that learned EEF action correction is decoupled from low-level execution, allowing the refiner to work with different robot controllers. On PiperX, we convert predicted corrections into EEF pose targets and use an IK solver to obtain joint commands. In LIBERO, interpolated EEF correction commands are executed by the environment's default operational space controller (OSC).

This modular design supports collision-aware execution with existing motion planners, as demonstrated by our cuRobo[5] integration. The EEF correction is converted into a target pose, toward which cuRobo plans a trajectory using current joint states, robot geometry, and scene obstacles. This integration improves task success without retraining the refiner, highlighting DreamChaser's modularity and extensibility (Table R6).

• Table R6. Correction performance and latency
ExecutionLIBERO-Plus SR (%)Added latency (ms)
FastWAM72.10.0
FastWAM + DreamChaser refiner76.811.3
FastWAM + DreamChaser refiner + cuRobo77.665.7

Q2. Performance of direct refiner control from predicted visual goals without WAM action chunks

Thank you for this constructive suggestion. We evaluate a video-to-control variant that discards WAM actions and uses the refiner to predict control targets from the current observation and predicted visual goal. On adapted LIBERO-Plus, this variant achieves 38.8% success, compared with 72.1% for FastWAM and 76.8% for DreamChaser (Table R7).

We attribute this difference to their complementary execution scales. The refiner specializes in goal-conditioned local control, learning short-horizon corrections toward predicted visual goals. The Action DiT jointly generates action chunks, capturing motion dependencies for sustained task execution. DreamChaser combines sequential action generation with targeted local correction to benefit from both.

• Table R7. Controller comparison
Controlleradapted LIBERO-Plus SR (%)
FastWAM actions only72.1
FastWAM video + DreamChaser refiner only38.8
FastWAM + DreamChaser refiner76.8

Q3. Choosing H and H_R and how their relationship affects performance

Thank you for this valuable question. H determines the WAM replanning interval, while H_R determines correction duration. We choose H_R = 4 as the shortest correction aligned with FastWAM's four-control-step video sampling interval, allowing collected trajectories to be reused in the original self-improvement training pipeline.

On adapted LIBERO-Plus, we compare H = 8, 16, 32 and H_R = 4, 8, 12 using the third-round WAM with the refiner enabled. The default H = 16, H_R = 4 achieves the highest success with shorter episodes than H = 8 (Table R8). Longer corrections reduce success and increase episode time, supporting brief corrections with a moderate WAM execution interval.

• Table R8. Horizons after three rounds on adapted LIBERO-Plus
HH_R=4 SR (%)H_R=4 Episode duration (s)H_R=8 SR (%)H_R=8 Episode duration (s)H_R=12 SR (%)H_R=12 Episode duration (s)
881.697.28————
1681.978.3481.280.2580.982.31
3279.672.0678.574.5677.976.82

Q4. Whether to freeze the refiner during WAM fine-tuning and the benefits of updating it

Thank you for this valuable suggestion. To reduce online training overhead, our original design trains the refiner once on the original demonstrations and freezes it during online improvement.

We also evaluate updating the refiner after each round of collection and WAM training, using expert replay and accepted online state-transition pairs. The settings share the first-round data and WAM checkpoint. Subsequent collections use their respective refiners. Fine-tuning the refiner further improves both WAM-only performance and execution with refinement (Table R9).

• Table R9. Refiner fine-tuning on adapted LIBERO-Plus
Refiner (SR, %)InitialInitial + refinerR1R1 + refinerR2R2 + refinerR3R3 + refiner
Frozen72.176.876.880.777.879.380.181.9
Fine-tuned each round72.176.876.881.278.379.780.582.4

Q5. Correcting the “color gitter” typo in Figure 3

Thank you for spotting this typo. We will correct “color gitter” to “color jitter” in Figure 3 and check the remaining figure labels to improve the clarity and accuracy of the manuscript.

References

[5] Balakumar Sundaralingam et al. CuRobo: Parallelized Collision-Free Robot Motion Generation. ICRA, 2023.

We are grateful for your thoughtful feedback and generous efforts in reviewing our work. Thank you for helping us improve the paper.

Sincerely,

The Authors

Dear Reviewer b9uT,

Thank you for the time and expertise you have generously devoted to reviewing our paper and supporting the research community.

Our responses include additional evaluations, baseline comparisons, and ablation studies, along with clarifications of the evaluation protocol, control interface, and computational efficiency. We hope these additions help address your concerns. When you have an opportunity, we would greatly appreciate your feedback and would be happy to provide any further clarification.

Thank you again for your valuable time and consideration.

Best regards,

The Authors

Reviewer_D4X1OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposed a self-improving framework for WAMs. Motivated by the video-action gap, this paper introduces a refiner, which use the video prediction to guide the action prediction. The refined chunks are further aggregated for fine-tuning the WAM.

Critical strengths

This paper shows a clear logical architecture, starting from the video-action gap and then proposing the trigger and refiner, finally aggregating the chunks for fine-tuning.

Critical weaknesses
  1. The video-action generalization gap, which serves as the core motivation of this paper, needs more analysis and evidence to verify its existence. I think it is not enough to only list some references. To better elaborate the idea of this paper, we need to know more about the gap.
  2. The improvements shown in Table 1 require justification because the refiner gain is marginal.
Questions for authors
  1. I understand that the refinement is an additional process, so what is the latency of refiner? I hope the authors can provide the computing time because realtime performance is also important for WAMs.
  2. (A minor suggestion) I think the figures could be polished to make the manuscript more readable.
3 comments

Dear Reviewer D4X1,

We appreciate your thoughtful assessment and valuable suggestions. Having carefully considered your comments, we present our detailed responses below.

W1. Evidence for the video-action generalization gap and its role in motivating DreamChaser

Thank you for highlighting this important point. WAMs inherit visual and motion priors from large-scale video pretraining, whereas action learning relies on much smaller paired robot datasets. During joint robot training, video prediction receives dense spatiotemporal supervision across many visual tokens, while action supervision targets low-dimensional control sequences. These differences in pretraining scale and supervision granularity are one motivation for studying the video-action gap.

We additionally collect 200 paired clips on adapted LIBERO-Plus and 200 across the three real-robot tasks: cup sleeving, plate placement, and bottle straightening. Each pair contains the WAM-generated video for an action chunk and the video recorded while executing that chunk's actions. Two human experts assess whether each clip shows reasonable task-directed behavior (Table R1). The gap is more pronounced on real robots, where only the predicted video is judged reasonable in 48.0% of pairs, compared with 13.0% in simulation.

• Table R1. Expert assessment of paired prediction and execution
SettingClip pairsBoth reasonable (%)Only video reasonable (%)Only execution reasonable (%)Neither reasonable (%)
adapted LIBERO-Plus20081.513.02.53.0
Real robot (3 tasks)20043.048.01.08.0

Meanwhile, we use Robometer[1], which is a pretrained robotic reward model that estimates task progress from videos and task instructions. We sample 1,000 video pairs from real-robot rollouts, pairing each WAM-generated clip with the video recorded during execution of its associated action chunk. The same frozen Robometer assigns mean progress scores of 0.546 to predicted videos and 0.192 to execution. Predicted progress is higher in 97.7% of pairs (Table R2). Together, these human and automated evaluations quantify the video-action gap and support visual guidance for local action correction.

• Table R2. Predicted versus executed progress estimated by Robometer
Clip pairsVideo progressExecuted progressVideo leads (%)
1,0000.5460.19297.7

W2. Justifying the marginal refiner gains in the main paper's Table 1

Thank you for this important comment. Most LIBERO tasks already approach 100% success, limiting the average improvement. Refinement improves performance on tasks that are not yet saturated. We illustrate these gains with the two tasks with the lowest baseline success, which improve from 88% and 90% to 94% each, raising their mean from 89% to 94% (Table R3).

We further evaluate all 24 tasks of the more challenging RoboCasa-GR1 benchmark. Three rounds of DreamChaser improve success from 67.8% to 72.0%, while vanilla fine-tuning declines to 62.2% (Table R4). These results support the framework's broad applicability and sustained improvement across diverse manipulation tasks and benchmarks.

• Table R3. Examples of refinement gains on LIBERO
Task (SR, %)FastWAMFastWAM + DreamChaser refiner
Wine bottle to cabinet top88.094.0
Black bowl: top drawer to plate90.094.0
Mean89.094.0
• Table R4. Full RoboCasa-GR1 evaluation
Method (SR, %)InitialInitial + refinerR1R1 + refinerR2R2 + refinerR3R3 + refiner
FastWAM + Vanilla-FT67.8—65.1—63.3—62.2—
FastWAM + DreamChaser67.868.569.670.271.071.572.072.4
Note: R1–R3: self-improvement rounds.

Q1. Additional refiner latency and its implications for real-time WAM execution

Thank you for highlighting runtime efficiency. On an RTX 5090, mismatch detection takes 9.7 ms on average (P95: 9.8 ms), and refiner inference takes 11.3 ms (P95: 11.6 ms), as reported in Table R5. Detection includes parallel encoding of the predicted and observed RGB endpoints and L2 computation. The refiner runs only when the mismatch trigger is activated.

• Table R5. Component latency
ComponentMean (ms)P95 (ms)
Mismatch detection9.79.8
Refiner inference11.311.6

Q2. Polishing the figures to improve the clarity and readability of the manuscript

We appreciate your thoughtful suggestion to improve the figure's clarity. We will carefully revise the overview figure to make the workflow and key components easier to follow.

References

[1] Anthony Liang et al. Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons. RSS, 2026.

We sincerely appreciate your insightful recommendations and careful attention to our paper. Your comments have been very helpful in strengthening the work.

Sincerely,

The Authors

Thank you for your response. I will keep my rating of weak acceptance.

Thank you for considering our clarifications and additional evidence, and for your continued positive assessment. We believe that your suggestions and the enhanced experimental evidence have significantly strengthened our work, which will contribute meaningfully to the community.

Reviewer_QPRWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a self-improving framework for World Action Models (WAMs). The hypothesis of the paper is that WAMs have a video-action capability gap where WAMs are able to predict plausible future visual outcomes but still fail to generate actions that achieve those outcomes. The paper discusses that these two capabilities are unequal because video models benefit from large-scale video pretraining, however learning robot actions requires comparatively scarce action-annotated data. To address this gap, the method presents an iterative self-improving loop: discrepancies between predicted and observed visual outcomes are calculated using VAE latent distances, a lightweight action refiner generates corrective actions conditioned on the predicted visual target, selective data aggregation retains correct trajectories, and WAM is finetuned on these collected trajectories along with the original expert demonstrations. Experiments are conducted on simulated benchmarks: LIBERO, LIBERO-plus, and real-world robotic manipulation tasks. The results show improvements in action execution and task performance. Overall, the contribution of the paper is proposing an iterative, self-improvement method to improve the action accuracy and task success rate by using the predicted WAM visual outcome to refine the actions.

Critical strengths
  1. A main strength of the proposed work is that it improves WAM through environment interactions and doesn't require additional human corrections or expert demonstrations during online rollouts. This has practical value because collecting expert demonstrations and corrective robot trajectories can be expensive.

  2. The paper shows evaluation across simulation benchmarks: LIBERO, an adapted LIBERO-plus benchmark, and real-world robotic manipulation tasks using AgileX PiperX. The real-world experiments are particularly valuable as they show that the method is not restricted to simulation.

Critical weaknesses
  1. The method primarily uses the predicted videos for action refinement. However, there is a main limitation that the predicted visual outcome may not be accurate, physically feasible, or achievable through the robot actions. A large discrepancy between the predicted video and actual execution may not necessarily indicate that the actions were incorrect. It can also happen from inaccurate video predictions itself. For example, if WAM predicts an unrealistic grasp/object manipulation, the refiner may be refined to guide the robot towards an incorrect target.

  2. To detect the video-action mismatch, the method uses the L2 distance between the VAE latents of WAM-predicted future frame and actual observation. However, the discrepancy in VAE latent space may not be equivalent to discrepancy in action space. VAE latent distance may be sensitive to appearance, object position, or camera-view changes rather than the accuracy of actions. It is possible that the correction may reduce visual discrepancy without improving the action accuracy or task completion.

  3. The experiments mainly show that using refiner helps in improving the task performance but it is still unclear whether the improvement specifically comes from using the WAM-generated future video to refine the action. The experiments do not isolate the contribution coming from using the predicted videos for refinement. If in a scenario, a separately trained corrective controller is used, would that also yield the same improvement?

Questions for authors

What is the motivation of using the final frame to find the discrepancy between WAM-predicted future frame and actual observation, over using the entire videos to calculate the discrepancy? Does using a single frame versus a video provide more benefits?

6 comments

Dear Reviewer QPRW,

Thank you for your careful review and insightful comments. We appreciate your constructive suggestions. After careful consideration of your suggestions, we provide the following detailed responses.

W1. Whether imperfect visual predictions can still guide effective action refinement

Thank you for this constructive comment. This is an important and insightful point. The motivation for video-guided action refinement is a capability gap in WAMs: they can often predict reasonable task outcomes, but their generated actions fail to realize them. An important source of this prediction-execution gap is that video prediction inherits visual and motion priors from pretraining on large-scale video data, whereas action learning relies on limited paired robot demonstrations. To further support this motivation, our evaluation covers 200 same-chunk prediction–execution pairs on adapted LIBERO-Plus and 200 across the three real-robot tasks. Assessments by two human experts show that predicted outcomes are more frequently judged reasonable than executed outcomes (Table R1). In addition, Robometer[1], a pretrained robotic reward model, compares task progress between WAM-predicted and actual execution clips of the same action chunk. Predicted progress is higher in 97.7% of 1,000 real-robot pairs (Table R2).

• Table R1. Expert assessment of paired prediction and execution
SettingClip pairsBoth reasonable (%)Only video reasonable (%)Only execution reasonable (%)Neither reasonable (%)
adapted LIBERO-Plus20081.513.02.53.0
Real robot (3 tasks)20043.048.01.08.0
• Table R2. Predicted versus executed progress estimated by Robometer
Clip pairsVideo progressExecuted progressVideo leads (%)
1,0000.5460.19297.7

As shown above, predicted videos indicate greater task progress than actual execution in the vast majority of evaluated real-robot pairs. Although video predictions are not always perfect, this does not negate DreamChaser's core contribution: using stronger visual priors for local correction and collecting video-action pairs for self-improvement. We appreciate the reviewer's emphasis on prediction reliability. To further strengthen robustness, DreamChaser's modular framework can integrate progress estimators such as Robometer[1]. Correction is triggered only when a visual mismatch is detected and the predicted clip receives a higher progress score than the execution clip. Otherwise, the WAM replans from the actual observation. This gate improves success with and without the refiner across three rounds (Table R3).

• Table R3. Progress-aware gating on adapted LIBERO-Plus
Gate (SR, %)InitialInitial + refinerR1R1 + refinerR2R2 + refinerR3R3 + refiner
Without72.176.876.880.777.879.380.181.9
Robometer progress gate72.177.377.480.978.479.680.582.3

More importantly, DreamChaser's self-improvement enhances not only action generation but also video prediction, providing more effective visual guidance for subsequent action refinement. As shown in Table R4, DreamChaser also improves video prediction using robot video-action pairs collected during interaction. In this experiment, the Action DiT and refiner remain fixed, and only the Video DiT is updated over three rounds. Both prediction–execution PSNR and task success improve across rounds, providing more reliable visual guidance for subsequent correction.

• Table R4. Video-only self-improvement on adapted LIBERO-Plus
RoundSuccess (%) ↑Prediction–execution PSNR (dB) ↑
Initial + refiner76.818.7
R1 + refiner78.221.9
R2 + refiner79.522.6
R3 + refiner80.223.1
Note: The Action DiT and refiner remain fixed across paired evaluations.

References

[1] Anthony Liang et al. Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons. RSS, 2026.

W2. Whether VAE latent discrepancy reflects action errors and whether reducing it improves task progress or completion

Thank you for this valuable question. We initially chose VAE features to reuse the WAM's existing modules and reduce additional computation. FastWAM already includes the VAE and generates future videos in its latent space, so no additional pretrained visual encoder is needed.

DreamChaser is a general framework that can integrate different visual encoders for discrepancy measurement, such as DINOv3, V-JEPA 2, and CLIP. Across all four evaluated representations, most corrections that reduce feature distance also improve task progress. Robometer progress increases in 82.8–96.1% of these cases, including 95.6% for VAE (Table R5). All four yield consistent self-improvement gains over three rounds. VAE provides the lowest detection latency while achieving comparable policy performance.

• Table R5. Correction-data selection and latency
SelectionInitialR1R2R3Progress gain when distance drops (%)Mean (ms)P95 (ms)
CLIP ViT-L/1472.176.577.779.182.812.814.5
DINOv3-L/1672.176.877.580.096.114.714.9
V-JEPA 272.177.077.980.196.148.351.6
VAE72.176.877.880.195.69.79.8
Note: Share of distance-reducing corrections that also increase Robometer progress.

W3. Whether the gains come from predicted-video guidance or could also be achieved by an independently trained corrective controller

Thank you for this constructive suggestion. We compare refiners with and without goal-image conditioning under the same experimental settings. Conditioning on WAM-predicted goal images improves correction success across all three rounds, with higher rates of goal-distance reduction and Robometer progress improvement (Table R6). These results demonstrate that the predicted goal image provides effective guidance for the refiner.

• Table R6. Target conditioning on adapted LIBERO-Plus
MethodInit.Init. + refinerR1R1 + refinerR2R2 + refinerR3R3 + refinerCloser to goal (%)Progress gain (%)
FastWAM + DreamChaser refiner (no future image)72.173.274.274.275.575.676.175.946.251.5
FastWAM + DreamChaser refiner (predicted future image)72.176.876.880.777.879.380.181.985.483.7

Q1. Why use only the final frame to detect prediction-execution discrepancy, and what benefits does it offer over full-video comparison?

Thank you for this question. The refiner intervenes after each WAM action chunk, so endpoint comparison focuses on the local execution outcome. We do not use the entire videos to calculate the discrepancy because predicted and executed sequences may progress at different rates, placing similar intermediate states at different times. Endpoint comparison reduces dependence on precise intermediate-frame synchronization.

On adapted LIBERO-Plus, endpoint comparison achieves higher success with and without the refiner across all three improvement rounds, while requiring less detection time (Table R7). These results support its effectiveness and efficiency for local correction and self-improvement.

• Table R7. Trigger comparison on adapted LIBERO-Plus
Trigger metric (SR, %)InitialInitial + refinerR1R1 + refinerR2R2 + refinerR3R3 + refinerDetection (ms)
Endpoint72.176.876.880.777.879.380.181.99.7
Multi-frame mean72.174.275.678.176.077.278.178.831.3

Thank you again for your valuable feedback, which has helped us improve the clarity and rigor of our work.

Sincerely,

The Authors

Dear Reviewer QPRW,

Thank you for generously devoting your time and expertise to reviewing our work and serving the research community.

Our responses provide additional evidence on visual-goal reliability and the relationship between visual discrepancy and task progress. We also include ablations on goal-image conditioning and endpoint versus multi-frame comparison. We hope these additions help address your concerns. When you have a chance, we would be grateful for your feedback and would be happy to clarify any remaining questions.

Thank you again for your time and consideration.

Best regards,

The Authors

Comments on the paper

1 comment

Dear Reviewers, 

We’re about halfway through the discussion period. Please read the authors’ rebuttal and assess whether it adequately addresses the key concerns raised in your review. We encourage you to leave a follow-up comment for the authors, indicating which concerns have been resolved and whether any important issues remain or require further clarification. Please update your review and scores as appropriate in light of the rebuttal and discussion.

Thank you for your time and participation. 

Best regards, 

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.