Deterministic World Models for Verification of Closed-loop Vision-based Systems
TLDR
Proposes a deterministic world model for verifying closed-loop vision-based systems, eliminating stochastic latent variables to achieve tighter reachable sets.
Reasoning
Strengths: Novel deterministic approach reduces overapproximation error and integrates conformal prediction for statistical bounds. Weaknesses: Limited to verification context, no mention of real-world deployment or scalability beyond benchmarks.
Read-first score
Read-first score 51.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 28.
Field roles
Rank sensitivity
Stability: volatile; rank range: 404.
Keyword Scores
Deep Analysis
Innovations
- Deterministic World Model (DWM) that maps system states directly to generative images, eliminating stochastic latent variables to reduce overapproximation error
- Dual-objective loss function combining pixel-level reconstruction accuracy with a control difference loss to maintain behavioral consistency with the real system
- Integration of DWM with Star-based reachability analysis (StarV) and conformal prediction to derive rigorous statistical bounds on trajectory deviation
Methodology
The DWM maps system states directly to generative images, trained with a dual-objective loss that balances pixel reconstruction and control difference. It is integrated into a verification pipeline using Star-based reachability analysis (StarV) and conformal prediction to derive statistical bounds on the trajectory deviation between the world model and the actual vision-based system.
Key Results
Experiments on standard benchmarks show that the approach yields significantly tighter reachable sets and better verification performance than a latent-variable baseline.