Optimization of humanoid walking controller: Crossing the reality gap

摘要

Humanoid locomotion remains a challenge because of the inherent instability of such robotic platforms. Inspired from observations on animals, Central Pattern Generators have been proposed to support the generation of rhythmic patterns able to make a robot smoothly walk while requiring few computational power. Nevertheless, tuning such controllers is challenging, in particular because small irregularities in the walking pattern easily make the robot fall. Optimization algorithms can be used to tune them in simulation, but the transfer of such solutions to the real robot raises the reality gap problem, as a solution efficient in simulation may well be inefficient in reality. It is proposed here to use the transferability approach to solve this problem. Its principle is to learn a model of the transferability between simulation and reality while doing several evaluations on the real robot. This model is then used to estimate how well a controller will transfer onto the real robot and the optimization process tries to optimize it besides other cost functions related to locomotion and tested in simulation only. The approach has been applied to the DARWIN-OP robot.