New AI controller enables lifelike robotic sprints in hours
AI learns human sprinting from scratch using just a 1-hour training session
A novel λ-hold controller, inspired by the equilibrium-point hypothesis from human motor control, lets a muscle-actuated skeletal model learn human-like sprinting using only a minimal task reward—and it does so within an hour of training. By holding each per-muscle threshold length over an interval of the gait phase, the approach sharply reduces how often the control policy must be queried, making exploration far more efficient in the body's high-dimensional, redundant action space. The method is grounded in physiology, connecting the equilibrium-point hypothesis, intermittent control, and optimal feedback control, and it offers a learnable model of the human motor controller.
- λ-hold controller uses equilibrium-point hypothesis to guide muscle excitation, reducing control space complexity
- System learned full-body sprinting in 1 hour with minimal reward vs. days/weeks for traditional RL approaches
- Reduces control frequency by 90% while maintaining human-like movement precision through physiological modeling
Why It Matters
This breakthrough could accelerate humanoid robot development by making human-like movement training exponentially faster and more efficient.