Gait Optimization on a Quadrupedal RHex using Multiarmed Bandits
Gait synthesis and optimization is a key challenge in the field of legged robotics. The performance of the system often relies heavily on the parameters that define a gait. Increasing the number of parameters allows a gait to be more finely tuned but in turn increases the difficulty of the gait parameter optimization problem due to the curse of dimensionality. The dependence on a physical machine makes it costly to measure the performance of a given gait, and inconsistencies in the model such as motor heating and battery placement make the optimization process susceptible to noise. In this paper we demonstrate a new method for tuning parameters based on formulating the gait optimization problem as a multi-armed bandit problem. The method is designed to account for uncertainty resulting from measurement noise while requiring a low number of physical trials. We tested this method on QRHex, a quadrupedal, RHex-style robot with one actuator per limb. With this method we were able to decrease the cost of transport by 36.6%.