Teaching Humanoid Robots to Move Like Us
Humanoid robots are getting steadily better at interacting with the world, but their movements still tend to be fairly clumsy. Teaching one to run, jump, or dance requires considerable work, often relying on controllers tuned for specific motions. Researchers from UC Berkeley and Stanford University are taking a different approach with BeyondMimic, an AI framework that teaches a humanoid robot a large collection of human-like movements from the same training setup.
The researchers demonstrated BeyondMimic on a Unitree G1 humanoid robot. Instead of manually programming each movement, they started with about 2.5 hours of human motion data covering everything from ordinary walking and running to dancing, martial arts, jumps, cartwheels, and other athletic movements. Those motions were adapted to account for the G1's different proportions and joint arrangement.
BeyondMimic teaches robots a wide range of behaviors (📷: Q. Liao et al.)
BeyondMimic then uses reinforcement learning to train a...
Copyright of this story solely belongs to www.hackster.io. To see the full text click HERE
