acceptodds
Withdrawn from ICLR 2027

AWARE: Action Weak-spot Assessment before Robotic-world-model Execution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Action-conditioned world models serve as environment surrogates for robotic planning, policy evaluation, and control, yet visually plausible predictions do not guarantee accurate modeling of action effects. Small action perturbations can induce response errors that propagate across successive prediction windows and ultimately produce substantial prediction drift. Existing visual and textual attacks conflate perceptual, semantic, and dynamics-prediction errors, making action-response errors difficult to isolate. We study the action-conditioning input as a separate attack surface and formalize pre-deployment action weak-spot discovery: identifying directions likely to induce large action-response mismatch between a world model and its environment without access to future environment responses. We propose AWARE (Action Weak-spot Assessment before Robotic-world-model Execution). Holding observation history and task semantics fixed, AWARE constructs a candidate pool under a shared budget in normalized action space and selects the candidate with the largest frozen-model response energy relative to the nominal prediction. Local response-geometry analysis characterizes the directional-alignment conditions and information limits under over-response and under-response. Compared with same-pool random selection, AWARE achieves higher mean true action-response mismatch at every tested budget: relative increases of – on CALVIN/Ctrl-World, – on RoboCasa/Ctrl-World, and – on CALVIN/IRA-SIM. AWARE provides a diagnostic of action–outcome consistency before deployment. Exploratory repair improves selected response measures but does not consistently reduce mismatch on a fixed episode-disjoint development set.

Primary area:
applications to robotics, autonomy, planning
Keywords:
action-conditioned world model, world model, attack, embodied AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WdoxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes AWARE, a method for selecting actions that may expose errors in a robotic world model before execution. With the visual history and task conditions held fixed, it generates 64 perturbations to the first action and selects the candidate that produces the largest change in the frozen model's prediction relative to the nominal action. The selected actions are then evaluated offline in a simulator by comparing the model's predicted action response with the environment's actual response. Across three environment–model combinations and four perturbation budgets, the paper reports higher average response mismatch than random selection from the same candidate pool. The authors also fine-tune the model using selected actions and simulator supervision. This repair improves some component metrics but does not reduce overall response error on held-out episodes.

Critical strengths

AWARE requires neither additional training nor access to future environment outcomes during selection. Across three environment–model combinations and four budgets, it selects actions with higher average response mismatch than random selection from the same pool. This is a positive empirical result and provides initial support for using the heuristic to select offline test cases.

Critical weaknesses
  1. The response-energy score is not sufficiently justified as a proxy for action-response mismatch. A large predicted response may be accurate, while an important action effect that the model fails to predict may receive a low score. Section 3.4 and Appendix A provide a conditional explanation, but the paper needs stronger evidence that these conditions hold in the tested settings.

  2. Beating random selection does not establish an advantage over simple alternatives. Candidates vary in magnitude and may be clipped to action bounds, so larger executed perturbations could produce both larger model responses and larger mismatch. Comparisons with selecting the largest executed perturbation, and with directions at the same executed magnitude, would help determine whether AWARE adds useful information beyond perturbation size.

  3. Task relevance and use on physical robots are not established. The paper does not show whether larger pixel-level action-response mismatch leads to poor action choices or task failure. Evaluation and supervised repair also rely on paired environment outcomes from the same initial state. How comparable responses would be obtained and used on physical robots remains unclear.

  4. Repair does not reduce overall response mismatch on episodes not used for repair. On the episode-disjoint development set in Table 3, response RMSE rises from 0.09833 to 0.10222–0.11001 across all three repair configurations. Improvements in direction or magnitude do not establish better overall response accuracy. The experiment therefore does not demonstrate that AWARE-selected samples enable effective repair across episodes.

  5. Figure 2 is difficult to read. The small labels and robot outlines are blurry. The figure should be redrawn or replaced with a clearer, higher-resolution version.

Questions for authors

Beyond higher average mismatch than random selection, what evidence shows that response energy identifies directions with large actual errors? Have the authors examined candidates with high response but low error, or low response but high error, within the same pool, including cases that AWARE misses? What evidence supports the directional alignment conditions in Section 3.4 in the actual evaluation settings?

Reviewer_Yd6SOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

In short, the authors provide a method which seeks to find perturbations that are likely to induce the most “surprise” in a model; i.e., how much does the model expect the perturbation to impact predictive performance, versus how much the perturbation affects performance in the ground-truth simulator. This allows them to determine where the model is likely to fail, and apply this knowledge downstream. Specifically, the authors applied this information for robotic manipulation tasks from CALVIN and RoboCasa, using the Ctrl-World and IRA-SIM world models and demonstrated that their perturbations induced larger errors than randomly selected perturbations.

Critical strengths
  1. The paper provides a way to understand the effect of perturbations without having to use the real-world (or simulator) at inference time. This makes the approach practical for deploying on hardware.
  2. The paper does provide some theoretical analysis and conditions under which their hypothesis (model-response sensitivity correlates with prediction mismatch) holds and evaluates this.
Critical weaknesses
  1. The novelty of the work is quite limited. The method can be described in Eq. 3 and is effectively searching for the perturbation that induces the largest error. Furthermore, they don’t provide any guarantees regarding how likely their sampling based approach is going to achieve the maximum disturbance result.

  2. Additionally, AWARE primarily identifies action perturbations that maximize the magnitude of the learned world model's predicted response. Under a local linear approximation, this amounts to identifying high-gain directions of the model's action-response Jacobian, effectively measuring predicted local control authority rather than model error. However, high predicted control authority does not necessarily imply high prediction error. A world model may accurately predict large action-induced changes while substantially underestimating other action effects. Although the authors theoretically characterize conditions under which sensitivity and prediction mismatch correlate, the experiments do not establish whether these conditions hold broadly or whether the proposed sensitivity-based heuristic offers advantages over existing alternatives. Thus, the empirical improvements provide limited evidence that AWARE generalizes beyond the experiments for world-model failure discovery.

  3. The exposition and presentation of the work can use substantial improvement (some examples include):

    • In Figure 1, (a) and (b) look identical and don’t communicate the difference between “nominal” and “perturbed”. Also, for c3, it is misleading for the region to be spherical as this may not be true.
  • Table 1 should either be replaced with a paragraph or have more details regarding features. The current phrases are hard to interpret. This could also be resolved by elaborating its caption.
  • For the theoretical analysis (section 3.4), defining the linear and quadratic terms. However, there is no clear explanation for how Eq. 5 arises from Eq. 4.
  • Several results are missing detailed explanations. E.g. In the paragraph titled “World-model response is not a direct differentiable surrogate for the final metric.”, the authors provide the numbers without giving any actionable explanations.
Questions for authors

Q1: Could the authors detail how they plan to address exposition improvements including those brought up in W4.

Q2: In reference to W2, could the authors describe why they find approach (using the local linearized approximation) target model error as opposed to searching for regions with higher control authority (a desired property). More experiments strengthening, this distinction would be useful.

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

Action-conditioned world models are increasingly used to train robot policies and to perform predictive control on complex dynamical systems. A common failure mode of current world models is their sensitivity to the action distribution. The paper proposes to use this sensitivity, measured as an action-response mismatch, to quantify how accurately a model predicts the consequences of an action. Building on this metric, AWARE identifies actions that are "weak spots" of the model. The authors support the method with a theoretical analysis of the conditions under which such weak spots can be found, and validate it empirically on two video world models. They also show that the identified weak spots can be used to repair the model.

Critical strengths
  • Current world models are often sensitive to the actions produced during policy training or planning, because these actions can drive them out of distribution, toward (state, action) pairs far from the training set. This often leads to hallucinated trajectories and incorrect behavior on downstream tasks. The problem studied is therefore central to the current limitations of world models.
  • The paper combines theoretical and empirical analyses that characterize when AWARE can identify candidate actions with a high action-response mismatch.
  • The repair experiment is interesting and promising, since the same mechanism could be used in an online learning setting to guide exploration of the state-action space.
Critical weaknesses
  • The paper does not study how well the action-response mismatch correlates with failures in planning tasks. An informative experiment would be to compute this quantity on trajectories planned with existing checkpoints (e.g., DINO-WM or LeWM) and check whether the mismatch is higher on failed plans. This is feasible because these world models operate in simulated environments where ground-truth states are available.
  • The evaluation covers only two diffusion-based video world models, and the training of the Ctrl-World checkpoints is not described. Since AWARE relies on the response of a stochastic model decoded to pixels, it is unclear whether the findings transfer to deterministic latent predictors (e.g., DINO-WM, LeWM, or other JEPA-style models). Evaluating at least one such model would strengthen the claim of generality.
Questions for authors
  • Can the action-response mismatch and the response magnitude, computed over a planned trajectory, predict task success or failure?
  • Could the AWARE response magnitude be used to regularize the planning objective during search (e.g., in control setups such as those of DINO-WM or LeWM)?

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

Related papers

Discussion (0)

Sign in to comment.