Probabilistic gradient ascent with applications to bipedal robotic locomotion

Bipedal robotic locomotion is an emerging field within the multi-billion dollar robotics industry, with global initiatives (such as RoboCup, FIRA and the DARPA Robotics Challenge) striving toward the development of robots able to complete complex physical tasks within a human-engineered environment. This paper details the redevelopment of an omni-directional walk engine for the DARwIn-OP, with an improved online optimisation framework developed for 13 of its internal parameters. Applying two well-known optimisation algorithms within this framework yields significant improvement in walk speed and stability. A new non-convex optimisation algorithm (Probabilistic Gradient Ascent) is derived from a reinforcement learning framework and applied to the same task, yielding an average speed improvement of 50.4% and setting a new maximum speed benchmark of 34.1 cm/s.