Natural Dynamic Controllers for Legged Robots

Miriam Zacksenhouse, Technion


Natural dynamic controllers aim at exploiting the natural dynamics of the system to perform the task rather than accurately following a prescribed trajectory. Pushing a child on a swing by applying a pulse of force once per cycle, when the swing approaches the person on the ground, clearly exploits the natural dynamics of the swing. In contrast, holding and moving the swing, as one would do when an infant is sitting on the swing, exemplifies trajectory control. In the context of legged locomotion, the former would correspond to pushing/throwing the leg once per cycle, by applying a pulse of torque at the hip, and letting the natural dynamics of the body and leg to propel the robot forward. Additional torque pulses may be applied at the hip and at the other joints to facilitate walking and enhance robustness.

The activation and termination of the torque pulses can be determined by either an internal phase variable, or specific events. The former implements a simple 2-level central pattern generator (CPG), consisting of a rhythmic generator (RG) that controls pattern formation (PF). Triggering torque pulses at specific events implements a reflex that may enhance robustness.

I will review our work on natural dynamic controllers for legged robots, including: (i) adaptation to a wide range of slopes with once per cycle feedback; (ii) stability and robustness analysis; (iii) enhancing robustness with compensation controllers; (iv) demonstration on a mono-pedal robot, (v) optimization using reinforcement learning with policy gradient; and (vi) advantages of 2-level compared to 1-level CPG.