humanoid locomotion
Generic and dynamic locomotion control for humanoid robots across discrete terrains, including whole-body soccer kicking.
We study how humanoid robots can learn generic, dynamic locomotion across discrete and uneven terrain. Our control architecture combines a reinforcement-learning policy for high-level decisions such as gait selection and step placement with model-predictive control (MPC) and whole-body impulse control (WBIC). This lets a robot choose between walking, jumping, and leaping from terrain height maps while retaining the stability and precision of model-based whole-body control.
The same emphasis on dynamic, whole-body behavior extends to humanoid soccer kicking. Using human biomechanics, trajectory optimization, and imitation learning, we develop kinodynamically feasible kicks for PresToe, a 25-DoF humanoid with actuated toes. In simulation, PresToe executes dynamic instep kicks that propel the ball at speeds above 11 m/s.
Videos
Generic and dynamic locomotion
Biomechanics-inspired soccer kicking
Learn more in the UMass Robotics humanoid robot control project.