Our approach to merging machine learning with engineered locomotion
In addition to being our co-founder and chief technology officer, Jonathan Hurst is also a robotics professor at Oregon State University. In the video below, Hurst describes how the early evolution of our robots hinged upon applying machine learning in an engineered system. Hurst breaks down the connections between designing machines, the biomechanics of locomotion, and merging our understanding of legged locomotion with machine learning.
He discusses how engineering for dynamic behavior is an element that is just as important as taking into consideration torque, speed, and kinetics, and how learned policies can be used as a control method for highly dynamic legged robots. Hurst also breaks down how we incorporated the use of clock-based periodic rewards, a reduced-order model, force profiles, and optimization tools to achieve legged locomotion in our robots.
Hurst provides some important tips for achieving legged locomotion:
Learn a gait library and use a reduced-order model for training and control
Learn a task space
Forget reference trajectories or models; instead, describe locomotion by cost function
Give a really simple trajectory
