Have you ever seen a robot performing a couple’s dance, like the waltz?
Surprisingly, the researchers at the University of California, San Diego have developed an AI system called Exbody2 which has made this possible.
The system, when used with humanoid robots, enables the robots to learn and replicate realistic body movements , such as walking in a straight line, side-stepping, running, couching, punching, squatting and even performing a waltz.
Researchers explained that the algorithm made humanoid robot movements stable and expressive which was achieved by combining a carefully selected dataset with separate systems for tracking body movement and controlling speed.
ExBody2 uses a Sim2Real technique, where robots are trained with simulated motion data that is then transferred to real world actions.
This helps the robots copy human movements in a realistic way.
The framework also creates a variety of useful training data that matches the robot’s ability.
It uses a teacher-student training method combined with reinforcement learning to ensure dependable performance in robot.
Moreover, a local tracking strategy is applied to enhance precision by separating motion tracking from speed control.
lastly, it includes a Conditional Variational Autoencoder (CVAE) that helps the robot to make smooth, continuous movement over extended periods.
