Károly Zsolnai-Fehér presents a virtual parkour system designed to overcome the tradeoff between imitation and adaptation. Controllers trained only to copy motion can look natural but fail when an obstacle moves, while goal-driven agents often solve a course with awkward or unrealistic movement.
The method trains one controller in two settings at once. One classroom teaches it to reproduce 19 short human-motion clips, while the other rewards progress through varied obstacle courses. A learned discriminator scores whether generated movement looks human and suits the surrounding obstacle, encouraging the agent to compose familiar skills in new situations.
The results include longer and previously unseen course layouts, but the system still has limitations. Success on long levels is about 40 percent, recovery motions can look unnatural, and the stronger motion tracking comes with some loss in task success. The work nevertheless shows how very limited human demonstrations can seed adaptable physical behavior.
Watch on YouTube
