acceptodds
Withdrawn from ICLR 2027

Calibrating Imagination by Decision Regret

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Action-conditioned world models can replace some camera updates, but visual prediction error alone does not measure the consequence of acting on an imagined observation. DREAM-CAL estimates policy-conditioned decision regret from paired continuations and selects a refresh horizon using a block-calibrated upper bound. Across 37 tasks at nominal 25% visual refresh, DREAM-CAL achieves 1,575/2,220 (70.95%) versus 1,392/2,220 (62.70%) for a visual-uncertainty head, a difference of 8.24 percentage points from the integer counts. In a separate segment audit, harmful-event AUROC is 0.790 versus 0.665, while the visual head retains lower expected calibration error. Gains vary by task family, and four of 37 task counts favor the visual head. We evaluate closed-loop success, event ranking, regret-bound coverage, distribution shift, and inference cost. Under dynamics shift, coverage with the original calibration offset falls from 87.92% to 71.25%, and mean decision computation takes 452 ms.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world models, active visual sensing, decision regret, conformal calibration, robotic manipulation, observation scheduling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_jTEUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DREAM-CAL, a decision-regret-based framework for determining when a robot should refresh visual observations instead of relying on world-model predictions. The method policy-conditioned decision regret from paired continuations and selects a refresh horizon using a block-calibrated upper bound. However, this paper lacks substantial experiments on broader comparisons against other baselines and open benchmarks, as well as suspicious real-world experiments.

Critical strengths

The paper addresses a problem in WAM and reports some improvements in its own evaluation set.

Critical weaknesses
  1. Lack of standard benchmarks and comprehensive comparisons.
    The paper evaluates 37 custom tasks (Section 5.1), without reporting results on established public manipulation benchmarks such as RoboTwin, CALVIN, or a reproducible LIBERO-based evaluation suite. More importantly, the experiments only compare against their internal ablation models, without direct comparisons to ANY existing published methods, making it difficult to evaluate the results.
  2. Insufficient evidence and documentation of real-robot experiments.
    Appendix I claims real-robot experiments on Franka, but provides no photographs, execution videos, or qualitative visualizations of the actual robot experiments. The hardware and evaluation protocols are also insufficiently documented for independent verification. Figure 11 only presents aggregate statistics, without demonstrating the experimental execution. Given the importance of real-world validation, the lack of supporting experimental evidence is a serious concern.
Questions for authors

See weakness

Reviewer_WyHrOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces a new method called DREAM-CAL that under a fixed budget of images that can be used by a robot to inform its decisions, chooses images such that they are likely to help the robot with the correct actions in comparison to taking every n-th image. For that the authors propose "decision regret" that measure the cost of the robot not taking a real images for k steps inside a world model. A prediction head is trained that approximates the 90th and 50th percentile, a probability of harm and a failure category. When the robot operates a rule based on these properties decides whether to execute an action-chunk blindly or to retrieve a new image. The evaluation shows that DREAM-CAL improves over a visual head.

Critical strengths

I find the motivation promising and the correct way forward. Including the policy's performance into the objective makes sense compared to pure visual prediction error or similar, since the tasks success depends directly on the policy, but may not be dependent on the visual error at all.

Critical weaknesses

Unfortunately, see several critical weaknesses:

  1. The writing is very hard to understand. While I find the introduction mostly good, the method section is hard to understand since many things are left unexplained (e.g. what is the contact cost? Is this a general approach or specific to some robot? How are the hyper parameters chosen? ...). Additionally, many parts of the method are introduced without proper motivation. The experiments section is cluttered with numbers without much interpretation. Overall, in my opinion the paper needs a major rewrite with less detail, but the major contributions highlighted and explained more in depth (so focus less on the "what"s and instead present the "why"s).
  2. The evaluation seem to not use any of the baselines cited in the related work, but only their own "visual head" that is hardly explained anywhere. This makes it hard to judge, whether the method actually produces meaningful improvements over existing methods.
  3. I do not see the point of the method if it has higher latency than just using all the image observations (which is the upper bound on the performance), if the cost of sensing is merely the computational time that is used to process the images.
Questions for authors
  1. What is the contact cost? Why is At,kA_{t,k} needed?
  2. How does the visual head work?
  3. What is the benefit of the method if it is slower than just using all the image observations? Is there some cost of sensing that motivates this work other than computational time?
  4. Does the method work with very stochastic policies or very stochastic environments?
  5. How does choosing M other than 4 influence the performance?
Reviewer_34wvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes DREAM-CAL, a method for deciding when a robot should acquire new visual observations rather than continue using world-model predictions. It defines a policy-conditioned, cost-augmented regret target through paired continuations with full visual feedback versus imagined observations. A learned head predicts regret and harmful-event probability, and block-level conformal calibration provides an upper bound used to select the next visual-refresh horizon. Evaluation includes 37 simulation tasks and four physical manipulation tasks.

Critical strengths

The paper makes a useful distinction between visual prediction quality and the consequences of acting on predicted observations. Paired continuations provide a concrete supervision mechanism for evaluating those consequences. The experiments examine closed-loop success, harmful-event ranking, calibration, distribution shift, and component ablations. The physical study provides additional evidence beyond simulation, and the paper explicitly acknowledges limitations in coverage, latency, and experimental comparability.

Critical weaknesses

The practical motivation could be clarified. The opening socket-insertion example explains why visual feedback becomes more important near contact, but it does not establish why that feedback should be withheld in the first place. Is there a sensing, communication, or computational constraint that prevents the robot from simply acquiring a new image when needed? Without specifying and evaluating such a constraint, it remains unclear when replacing real observations with generated images is preferable, particularly given the method’s reported 452 ms mean computation time.

The experiments do not demonstrate that the cost saved by reducing visual updates outweighs world-model generation and risk-evaluation costs. The reported 452 ms mean computation time exceeds the nominal 100 ms control period. In the physical study, realized visual-refresh ratios are neither reported nor matched, making the improvement from 61/100 to 66/100 successes difficult to attribute specifically to better refresh scheduling.

Questions for authors

(1) What specific sensing, communication, or processing constraint makes imagined observations preferable to acquiring real images in the evaluated systems, and can the authors quantify the complete cost tradeoff? (2) Could the physical experiments report actual image counts and compare methods at matched visual-refresh budgets, including full-feedback task duration? (3) How does the block-level coverage guarantee relate to online execution when the refresh controller changes the visited-state distribution? (4) Since the regret target includes absolute contact cost rather than the difference between branches, how sensitive are the results to this choice, and would a reference-subtracted cost better isolate the consequences of withholding vision?

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

Related papers

Discussion (0)

Sign in to comment.