A New Control System Unifies Robot Walking, Running, and Jumping
Robots can put on some really impressive shows on stage when the conditions are controlled and everything is carefully choreographed. But things aren’t so easy in the real world. After stepping off the stage, they have to recognize when the terrain changes, decide whether to climb, jump, or slow down, and transition between those movements. Researchers at KAIST believe they have taken a significant step toward that goal with a new control framework called APT-RL, which allows a quadruped robot to autonomously select the most appropriate gait and switch between different locomotion skills, relying only on its onboard sensors.
The system, whose name stands for Action Pretrained Transformer-based Reinforcement Learning, replaces the traditional approach of using separate controllers for different movements with a single policy that can adapt continuously as the environment changes. Instead of treating walking, running, jumping, and obstacle traversal as independent tasks, the controller learns when...
Copyright of this story solely belongs to hackster.io. To see the full text click HERE
