Reactive Motion Generation via Phase-varying Neural Potential Functions

A team of researchers from EPFL, University College London, and KTH Royal Institute of Technology has developed Phase-varying Neural Potential Functions (PNPF), a novel framework for learning robot motions from demonstrations. Published in IEEE Robotics and Automation Letters, PNPF addresses a persistent challenge in robotics: generating stable, reactive motions for tasks where trajectories intersect, such as drawing a figure-8 or complex assembly sequences. Traditional dynamical systems (DS) methods struggle with these scenarios because first-order models require a unique velocity for each state, while second-order models using velocity to disambiguate direction become brittle near intersections. Phase-based methods, which rely on open-loop time, fail to recover after perturbations.

PNPF solves this by conditioning a learned potential function on a phase variable that is dynamically estimated from the robot's state progression, rather than from a fixed temporal schedule. This allows the system to handle state revisits naturally—the robot knows where it is in a task based on its actual progress, not just time. The learned potential function generates local vector fields that provide both stability and reactivity, enabling the robot to recover from external disturbances (like a human pushing its arm) without needing explicit replanning. In experiments, PNPF generalized effectively across point-to-point, periodic, and full 6D motion tasks, and significantly outperformed existing baselines on trajectories with intersections, demonstrating real-time robotic manipulation under real-world disturbances.