Upcoming Event: Oden Institute Seminar

Reliable World Models: Physical Interpretability for Prediction and Verification

Ivan Ruchkin,

Abstract

Generative neural architectures, known as world models, perform well in predicting action-conditioned raw observations, which is valuable for controller training and other engineering tasks. Unfortunately, learned latent representations in world models lack direct mapping to meaningful physical quantities, limiting their utility and interpretability in downstream planning, control, and safety verification. This talk starts by arguing for a fundamental shift from physically informed to physically interpretable world models. It presents four principles that leverage symbolic physical knowledge for interpretable world models: (1) structuring latent spaces, (2) aligning with invariances/equivariances, (3) exploiting supervision with varied strength and granularity, and (4) partitioning generative outputs.  The second part of this talk dives into an interpretable world model for trajectory prediction. We discuss a novel architecture that aligns learned latent representations with real-world physical quantities by combining physically interpretable image encoding and a partially known dynamical model. The architecture integrates a vector-quantized visual encoder, a transformer-based physical encoder, and a learnable dynamics layer with the known physical equations. To eliminate the impractical reliance on precise ground-truth physical knowledge, the training incorporates weak distributional supervision. In three case studies, we demonstrate that this architecture not only provides physical interpretability but also achieves superior state prediction accuracy.  The final part of this talk presents a Deterministic World Model (DWM) that enables reachability verification of vision-based systems. Specifically, we design a latent-free neural decoder that maps physical states (e.g., position and velocity) directly to synthetic camera images, removing the overapproximation error caused by stochastic latent variables. The DWM is trained with a dual loss combining saliency-map reconstruction and control consistency that preserves similarity to the real controller's behavior. We integrate the DWM into closed-loop reachability analysis and apply conformal prediction to inflate the reachable sets by a distribution-free trajectory-tube deviation bound, transferring the surrogate guarantee to the real system with high probability. Experiments show that the DWM produces tighter reachable tubes than the baselines of cGAN and neural predictor baseline and trajectory predictor baselines while meeting the target coverage after conformal inflation.

Biography

Dr. Ivan Ruchkin is a Malachowsky Family Endowed Rising Star Assistant Professor in the Department of Electrical & Computer Engineering at the University of Florida.  His research improves the safety, reliability, interpretability, and trustworthiness of autonomous systems by advancing their modeling, analysis, verification, monitoring, and prediction.  Previously, Ivan was a postdoctoral researcher at the PRECISE Center at the University of Pennsylvania. He received his PhD in Software Engineering from Carnegie Mellon University (CMU) and a Specialist Degree in Applied Mathematics and Computer Science (with honors) from the Lomonosov Moscow State University. Ivan also held several positions in the industry and government, including at the Air Force Research Lab (AFRL) and NASA Jet Propulsion Lab (JPL).  Ivan’s research on assuring cyber-physical systems has been recognized with the NSF CAREER Award and the Frank Anger Memorial Award for the crossover of ideas between the SIGSOFT and SIGBED communities. Ivan’s work received multiple Best Paper/Poster/Demo Awards as well as a Gold Medal in the ACM SRC student competition at the MODELS conference.