acceptodds
Under review as a conference paper at ICLR 2027

ActForesight: Predictive Safety for Embodied Agents with Action-Conditioned Visual Futures and Motion Monitoring

Abstract

Embodied agents must continuously perceive, reason, and act in dynamic physical environments, where the consequences of an action may unfold beyond the current observation. We present ActForesight, a closed-loop predictive safety framework that combines action-conditioned visual futures with independent motion monitoring for pre-execution physical-safety assessment. Given a candidate action, an action-conditioned world model generates multiple possible visual futures, which are analyzed by a shared Safety Critic to identify evidence of potential physical hazards. A risk-sensitive Any-Block aggregation rule converts future-level judgments into an action-level safety decision, blocking an action whenever any predicted future reveals a safety-critical consequence. In parallel, an independent motion-safety monitor evaluates candidate trajectories against explicit kinematic and geometric constraints, including joint limits, workspace boundaries, and geometric reachability. The two branches are integrated as a complementary safety filter before task-utility-based action selection, preserving the original task objective while incorporating predictive safety checks. Experiments in simulated robotic manipulation tasks show that action-conditioned visual futures capture action-dependent physical consequences, multi-future prediction broadens safety-critical consequence coverage, and motion monitoring complements visual reasoning by identifying constraint-related hazards. Together, these components provide a more comprehensive basis for pre-execution physical-safety assessment and demonstrate how world models and explicit motion constraints can jointly support safer embodied decision-making.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied intelligence, World model, Action-conditioned future prediction, Predictive security decision-making, Multi-future imagination, Robot safety

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_Ddq7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a safety filter for robotic manipulation. The core idea is simple: the predicted action is fed into an action condition world model which generates several visual futures. These are then fed into a classifier which determines if it represents a safe or unsafe rollout. In addition a more classic based safety filter is included which in parallel checks for joint and workspace violations. The filter is evaluated offline in simulation on several tasks. The world model future based safety filter catches a relatively small amount of safety violations, while the classical safety filter catches the majority of the safety violations.

Critical strengths

Robotic safety is an important area of research. This paper tackles a difficult problem of predicting failures before they happen using world models. This is a task that humans often do during manipulation and a robust system that can predict the effects of its actions on the environment and change them accordingly is important.

Critical weaknesses

Experiments are only conducted in simulation on a small number of tasks. Similar related work conducts experiments in the real world Uncertainty-aware Latent Safety Filters for Avoiding Out-of-Distribution Failures (Seo 2026), FOREWARN (Wu 2025).

I feel this paper generally offers limited novelty over prior work. Several papers are conceptually very similar including Latent Safety Filters for Avoiding Out-of-Distribution Failures (Seo 2026) (which is not cited) and FOREWARN (Wu 2025). At the very least prior related work which solves a similar problem should be included as baselines to the method. Additionally I feel it would make the paper stronger to also include linear probes on activations to detect unsafe behavior as a baseline. I find the motion safety monitor aspect of this paper to be a relatively minor contribution.

Examining the results in Table 2. It appears that the WM proposed method has a recall of only 4.67% of unsafe behavior at three predicted futures and 7.33% at five futures, while the motion safety monitor has a recall of 54.95% of unsafe behavior. This to me indicates that the proposed world model method does little to detect unsafe behavior.

Questions for authors

What is the nature of the world model ? Is it fine-tuned from an existing world model or simply running forward the environment you are using on several action generations ? As far as I can tell the world model being used is never explained.

How do you differentiate yourself from prior work like Uncertainty-aware Latent Safety Filters for Avoiding Out-of-Distribution Failures (Seo 2026) and why is it not included as a benchmark ?

Reviewer_o4R2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ActForesight, a safety-filtering framework for embodied agents that combines action-conditioned future prediction with a motion-safety monitor. Candidate actions are screened for potential hazards before task-based action selection, with the goal of improving physical safety without changing the underlying planner.

The idea is well motivated, particularly the use of multiple predicted futures and the risk-sensitive Any-Block aggregation strategy. The evaluation is also transparent, with thorough dataset auditing and detailed failure analysis.

However, the effectiveness of the core visual-future component is limited. Even the best configuration detects only 20 of 273 unsafe cases (7.33% recall), while the motion monitor accounts for most of the overall performance gains. In addition, all experiments are offline, so it remains unclear how much benefit the framework would provide in a closed-loop robotic system.

Critical strengths

The main strength of the paper is its explicit separation of physical safety from task success. Rather than modifying the underlying task objective, ActForesight treats safety as a pre-execution admissibility filter that removes hazardous candidates before utility-based action selection. This framing is conceptually clean and keeps the safety mechanism independent of the upstream planner.

The proposed Any-Block aggregation rule is well motivated for safety-critical settings. The experiments show that increasing the number of predicted futures improves hazard coverage, with unsafe recall increasing from 2.20% using a single future to 7.33% using five futures under Any-Block aggregation. In contrast, majority voting reduces recall to 1.47%, supporting the authors’ argument that low-frequency but potentially costly hazards should not be suppressed through consensus-based decisions.

The paper is also commendably transparent in its evaluation. The dataset audit clearly distinguishes verified SAFE/UNSAFE samples from unresolved cases, with only 594 of the 600 collected samples included in the final evaluation. The authors provide confusion counts, hazard-specific analyses, oracle diagnostics, and misattribution studies, making it easier to understand the sources of failure rather than relying solely on aggregate metrics.

Critical weaknesses

The main limitation is the weak performance of the visual-future branch. Even with five predicted futures, unsafe recall is only 7.33%, corresponding to 20 detected unsafe samples out of 273. More importantly, the executed-future oracle reaches only 6.37% recall, suggesting that the current critic interface struggles to identify most hazards even when provided with perfect post-action visual evidence.

The paper provides limited implementation detail for several key components. Critical aspects such as the architectures, model backbones, training procedures, prompting strategies, and computational cost of the world model and Safety Critic are not sufficiently described, making it difficult to assess reproducibility and practical deployment requirements.

Although ActForesight is presented as a pre-execution filtering framework for embodied agents, all evaluations are conducted offline on pre-collected state-action pairs. The paper does not demonstrate how the filtering mechanism affects task completion, replanning behaviour, stop decisions, or overall system performance in a closed-loop execution setting.

Finally, the benefits of fusion are driven largely by the motion monitor's ability to detect workspace-boundary violations. Hazard-specific analysis shows 98.68% recall for workspace violations but 0% recall for collisions, drops, topples, and joint-limit violations. As a result, fusion increases unsafe recall only modestly (54.95% to 57.88%) while introducing additional false blocks and slightly reducing the overall unsafe F1 score (70.92% to 70.38%). This raises questions about how broadly applicable the proposed safety framework is beyond the specific hazard categories covered well by the monitor.

Questions for authors

Despite the independent motion-safety monitor's explicit checks of joint limits through inverse kinematics, it records a recall rate of 0.00% for all 36 occurrences of joint-limit hazards in the dataset. This raises questions about the effectiveness of the monitoring system in identifying these hazards. Is this entirely due to the 69 missing planned joint trajectories and inverse-kinematics solver failures mapped to allow?

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

The paper presents ActForesight, a pre-execution safety framework for embodied agents. The method evaluates candidate robot actions before execution through two complementary branches. The first branch uses an action-conditioned world model to generate multiple possible visual futures. A shared Safety Critic assesses each future, and an Any-Block rule rejects the action if any predicted future provides evidence of a physical hazard. The second branch independently checks the planned motion against explicit kinematic and geometric constraints, including joint limits, workspace boundaries, and geometric reachability. Only actions accepted by both branches are passed to the original task-utility function. The framework is evaluated in RLBench/CoppeliaSim on four manipulation tasks, with 594 verified SAFE/UNSAFE samples. The experiments compare current-observation reasoning, one to five predicted futures, majority voting, an independent motion monitor, and their fusion. Increasing the number of predicted futures improves unsafe recall from 2.20% to 7.33%, while the motion monitor alone achieves 54.95% recall. The fused system reaches 57.88% unsafe recall with a 5.61% false-block rate.

Critical strengths

S1. The paper addresses an important problem with a clean modular formulation. Pre-execution safety is highly relevant for embodied agents. I like the separation between task utility and safety admissibility. An unsafe action cannot be compensated by a high task score, and the safety layer can in principle be added to different upstream planners.

S2. The evaluation is transparent and unusually careful about what the results do and do not show. The paper distinguishes task failure from physical danger, keeps unresolved cases visible in the audit, and reports several diagnostics rather than relying on a single aggregate metric. The authors also explicitly acknowledge that generated futures are neither calibrated probabilities nor guarantees of safety.

I think this careful treatment of evaluation is a genuine strength of the work.

Critical weaknesses

W1. The main predictive-safety branch detects only a very small fraction of unsafe actions. This is the central issue for me. With five predicted futures and Any-Block aggregation, unsafe recall is only 7.33%, corresponding to 20 detected unsafe samples and 253 missed unsafe samples. The method therefore allows more than 90% of the verified unsafe cases to pass through the visual safety branch. Increasing the imagination budget does improve recall relative to a single future, but the absolute performance remains very low. Collision recall is 0%, drop recall is 7.5%, topple recall is 6.25%, and workspace violations reach only 7.89% through the visual branch. This makes the current evidence for the paper's main novelty rather weak. The proposed use of visual futures is conceptually interesting, but I would expect a safety mechanism to demonstrate substantially stronger hazard discrimination before considering it a convincing contribution.

W2. Most of the final performance comes from the conventional motion monitor rather than from visual foresight. The independent monitor alone reaches 54.95% unsafe recall with zero false blocks. Fusion improves recall only to 57.88%, while introducing a 5.61% false-block rate. It also slightly decreases Unsafe F1 from 70.92% to 70.38%. The task-wise results make this even clearer. On push buttons and slide-block tasks, the monitor alone obtains 100% recall and the fused method adds nothing. On insertion and stacking, the monitor contributes nothing, and performance remains at only 8.33% and 5.75% recall respectively. It seems to me that the current system is mainly successful when an unsafe action violates an explicit workspace constraint. This is useful, but it is already handled by the simpler deterministic branch. The genuinely difficult hazards that motivate predictive visual reasoning, such as drops, toppling, and collisions, remain largely unresolved. A related concern is that even using actual post-action observations does not dramatically improve the Safety Critic. The oracle diagnostic reaches only 6.37% recall with key-frame sequences. This suggests that the bottleneck is not only the quality of the predicted futures, but also the ability of the current visual interface and critic to recognize physical safety events.

Questions for authors

I like the motivation and the modular formulation, and I think the paper is careful in its evaluation and in limiting its claims. Using predicted visual futures as pre-execution safety evidence is also an interesting research direction. However, the current results do not yet convince me that the proposed predictive component is effective enough. The five-future visual branch detects only 7.33% of unsafe cases. Most of the final performance comes from an explicit motion monitor, which is simpler and already achieves 54.95% recall with no false blocks. Fusion gives only a modest recall improvement and introduces additional false blocks. The diagnostic results also suggest that the problem is deeper than imperfect world-model predictions, since the Safety Critic performs poorly even when given actual future observations. For me, this weakens the main technical message of the paper. A stronger temporal safety representation, better recognition of object-level hazards, or a demonstration on a broader set of hazards where visual foresight provides a clear advantage over explicit constraints would significantly improve the work.

Q1. Can the authors explain more clearly what evidence supports the usefulness of the visual-future branch beyond the explicit motion monitor? In the current results, the monitor alone provides most of the final recall and even has slightly better Unsafe F1 than the fused system. I would be interested in an analysis restricted to hazards that cannot be captured by explicit geometric constraints.

Q2. The executed-future oracle also has very low recall. Does this imply that the main bottleneck is the Safety Critic rather than the world model? If so, have the authors tested a stronger temporal critic or a critic that receives a short video instead of individual frames?

Q3. Why does the runtime policy map UNKNOWN to ALLOW? From a safety perspective, unresolved evidence seems particularly important. Could the authors report the safety and utility trade-off obtained by treating UNKNOWN as BLOCK, or by introducing a third action such as stop/replan? The current rule may systematically favor execution in precisely those cases where the safety evidence is insufficient.

1 comment
Reviewer_UQNbOpenReview ↗

Is there any rebuttal for this paper?

I think it would be nice to discuss in any case this paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.