Trajectory Clustering for Robot Physical Intelligence

This project explores how robots can develop physical intelligence—the ability to learn and generalize motion patterns from human demonstrations. Inspired by crowdsourced learning and geographic information systems, it applies the TRACLUS algorithm to segment and cluster motion trajectories into meaningful sub-trajectories. By representing motion data (position, velocity, acceleration, and force) as high-dimensional trajectories in state space, the system identifies shared structural patterns across repeated demonstrations. Through optimization and clustering, it filters noise and groups similar motion segments, enabling robots to recognize and reproduce valid movement behaviors even from unseen trajectories. The work bridges robot learning from demonstration, trajectory analysis, and data-driven motion intelligence, offering a scalable framework for teaching robots through collective human motion data—an early step toward autonomous physical learning inspired by Westworld and The Three-Body Problem.