An AI-Powered Control System For Robots With Legs

Researchers led by Yilun Du have developed an AI-based robotic control system that allows four-legged robots to adapt their movements without retraining.
The new method, called Diffusion-MPC, blends model predictive control with reinforcement learning principles to optimize performance.
The Diffusion-MPC system uses a generative diffusion model instead of a fixed, scratch-built dynamics model to adapt its plans to new environments and tasks in real time.
The flexible control system demonstrated stability and adaptability in real legged robots, effectively handling terrain changes, slopes, and dynamic control objectives.
This advancement represents progress towards creating general-purpose locomotion controllers that can dynamically adjust to field conditions instead of relying on pre-set training scenarios.

You may have missed