What You Don't Know Can Hurt You: How Well do Latent Safety Filters Understand Partially Observable Safety Constraints?

arXiv 25.10 2025 34.4 theory, application

TLDR

Identifies failure modes in latent world models for safe control under partial observability, proposes diagnostics and mitigations, validated on robotic cooking tasks.

Reasoning

Strengths: novel identification of estimation and prediction gaps, practical diagnostics, real-world hardware validation. Weaknesses: mitigations increase conservativeness, limited to specific cooking tasks.

Read-first score

Read-first score 34.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 32.

Recency 6%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=experiment

Topical relevance 29%

45.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 142.

Keyword Scores

world model

world dynamics prediction

generative world model

model-based reinforcement learning world model

world simulator

interactive world model

video world model

Deep Analysis

Innovations

Identification of two distinct failure modes in latent world models under partial observability: estimation gaps (safety-relevant information not present in current observations) and prediction gaps (failures observable but not reliably anticipatable from available observations).
Introduction of two diagnostics: a mutual-information-based measure of safety observability and a rollout-based measure of future safety predictability.
Proposal of two mitigation strategies: privileged multimodal supervision for estimation gaps and conformal risk calibration for prediction gaps.

Methodology

The study analyzes latent-space safe control problems using world models trained from high-dimensional observations. It defines two failure modes—estimation gaps and prediction gaps—and proposes corresponding diagnostics and mitigation strategies. Experiments are conducted on a Franka Research 3 manipulator performing cooking tasks, comparing unimodal RGB world models with multimodal RGB+Tactile and RGB+Thermal variants.

Key Results

The proposed mitigation strategies improve safety of the robot manipulator under partial observability, but with increased conservativeness. The work demonstrates that partial observability can induce control failures when safety-relevant information is not preserved in the latent state.

Limitations

The mitigation strategies lead to increased conservativeness, which may limit task efficiency.
The study is limited to specific hardware (Franka Research 3 manipulator) and cooking tasks, so generalizability to other domains is not established.
The paper does not claim to fully resolve partial observability; it only raises the question of when world model state representations are sufficient for reliable robot control.

Resources

Paper