V-MORALS: Visual Morse Graph-Aided Estimation of Regions of Attraction in a Learned Latent Space

A research team from Carnegie Mellon University and UC Berkeley has introduced V-MORALS (Visual Morse Graph-Aided Estimation of Regions of Attraction in a Learned Latent Space), a novel approach to robot safety analysis that operates purely on visual sensor data. The system addresses critical limitations of existing reachability analysis methods, which typically require precise knowledge of system dynamics or extensive datasets to estimate accurate models. V-MORALS instead learns a low-dimensional latent space directly from image-based trajectories of a system under a given controller, then uses topological tools to generate well-defined Morse Graphs from which it computes Regions of Attraction (ROAs). This represents a significant advancement over the original MORALS method, which still relied on full state knowledge.

Technically, V-MORALS takes in datasets of image trajectories and learns a latent representation suitable for reachability analysis, enabling safety verification without access to the robot's internal state information. The method provides capabilities similar to the original MORALS architecture while eliminating the dependency on state knowledge and using only high-level sensor data. This breakthrough makes safety analysis applicable to real-world robotic systems where internal states are often inaccessible or poorly modeled, potentially accelerating deployment of autonomous systems in unstructured environments. The researchers have made their project website publicly available, suggesting this could become a foundational tool for verifying safety in vision-based robotic control systems.