HALO: Hybrid Auto-encoded Locomotion with Learned Latent Dynamics, Poincar\'e Maps, and Regions of Attraction

A team from Caltech and UT Austin has introduced HALO (Hybrid Auto-encoded Locomotion), a novel AI framework designed to solve a core challenge in robotics: guaranteeing stability for complex, legged machines. Traditional methods rely on simplified, hand-designed models that often fail to capture real-world dynamics, while purely data-driven approaches lack formal safety guarantees. HALO bridges this gap by using an autoencoder—a type of neural network—to automatically discover a compact, low-dimensional representation, or latent space, from high-dimensional robot motion data. Within this latent space, the system learns a Poincaré map, a mathematical tool that models the robot's step-to-step evolution, enabling rigorous stability analysis.

The key innovation is that HALO can construct a mathematically provable 'region of attraction' within this learned latent model. This region defines all the states from which the robot's locomotion will remain stable and recover to its intended gait. Crucially, the framework's decoder can then map these stability guarantees back to the robot's actual, full-dimensional state space. In experiments, HALO was validated on simulated systems including a hopping robot and a full-body humanoid, where it successfully predicted the boundaries where locomotion would fail. This means engineers can now use data to build models that not only predict robot behavior but also formally certify when a walking or running robot is on the verge of falling, enabling safer and more robust autonomous operation.