An AI-powered control system for robots with legs

Legged robots need to travel safely through rough and changing environments, but common control approaches have limitations. Model predictive control can optimize behavior, yet it depends on accurate models of robot dynamics that are difficult to obtain in real settings and often require simplifying assumptions. Model-free reinforcement learning can produce reliable behaviors, but those behaviors are generally fixed and hard to modify after training.

A Harvard-led team headed by computer science professor Yilun Du developed Diffusion-MPC, a control system intended to combine features of both approaches. The system uses a generative diffusion model, a type of AI commonly associated with image generation, together with an approximate model of how the robot and its surroundings will change over time. Rather than relying on a fixed, manually constructed model, the method uses this approximation to help plan motion.

Diffusion-MPC refines plans in real time through rewards for desirable performance and constraints covering physical limits and safety rules. Incorporating rewards and constraints during operation enables the robot to alter its behavior for new demands without additional training. According to the researchers, the system can adjust speed, direction, or gait as objectives change.

The team tested the research system on real quadruped robots. The demonstrations showed stable locomotion while the robots adapted to different objectives and obstacles. Reported scenarios included changing terrain, slopes, balance recovery, and different control objectives. Du described the work as a step toward general-purpose locomotion controllers that could be reprogrammed or adjusted in the field as conditions change, instead of being confined to the specific situations used in training. The work was presented at the 2026 IEEE International Conference on Robotics and Automation, and the associated paper is available on arXiv.