BiRRTOpt: A combined sampling and optimizing motion planner for humanoid robots

Abstract

Currently, the optimization-based methods are widely adopted on humanoid robots and other bipedal robots. However, using these methods to plan robot motions suffers from getting stuck in infeasible solution given bad initial guesses. Conversely, sampling-based methods which are probabilistically complete perform well in practice. But trajectories generated by these methods require smoothing techniques, normally in low-quality in terms of a given cost function. This paper proposes a software framework called BiRRTOpt which combines the advantages of both approaches, aiming to compute a feasible, collision-free and optimized trajectory for humanoid robots. This work is mainly divided into 2 phases: Initial Guess and Optimization. Two experiments in the simulator and on an actual Atlas robot demonstrate the success of our approach. The idea described in this paper can easily be extended to other humanoid robots or bipedal robots to plan motions.