Recent advances in deep reinforcement learning (RL) based techniques combined with training in simulation have offered a new approach to developing robust controllers for legged robots. However, the application of such approaches to real hardware has largely been limited to quadrupedal robots with direct-drive actuators and light-weight bipedal robots with low gear-ratio transmission systems. Application to real, life-sized humanoid robots has been less common arguably due to a large sim2real gap. In this paper, we present an approach for effectively overcoming the sim2real gap issue for humanoid robots arising from inaccurate torque-tracking at the actuator level. Our key idea is to utilize the current feedback from the actuators on the real robot, after training the policy in a simulation environment artificially degraded with poor torque-tracking. Our approach successfully trains a unified, end-to-end policy in simulation that can be deployed on a real HRP-5P humanoid robot to achieve bipedal locomotion. Through ablations, we also show that a feedforward policy architecture combined with targeted dynamics randomization is sufficient for zero-shot sim2real success, thus eliminating the need for computationally expensive, memory-based network architectures. Finally, we validate the robustness of the proposed RL policy by comparing its performance against a conventional model-based controller for walking on uneven terrain with the real robot.

Model-free or learning-based control, in particular, reinforcement learning (RL), is expected to be applied for complex robotic tasks. Traditional RL requires a policy to be optimized is state-dependent, that means, the policy is a kind of feedback (FB) controllers. Due to the necessity of correct state observation in such a FB controller, it is sensitive to sensing failures. To alleviate this drawback of the FB controllers, feedback error learning integrates one of them with a feedforward (FF) controller. RL can be improved by dealing with the FB/FF policies, but to the best of our knowledge, a methodology for learning them in a unified manner has not been developed. In this paper, we propose a new optimization problem for optimizing both the FB/FF policies simultaneously. Inspired by control as inference, the optimization problem considers minimization/maximization of divergences between trajectory, predicted by the composed policy and a stochastic dynamics model, and optimal/non-optimal trajectories. By approximating the stochastic dynamics model using variational method, we naturally derive a regularization between the FB/FF policies. In numerical simulations and a robot experiment, we verified that the proposed method can stably optimize the composed policy even with the different learning law from the traditional RL. In addition, we demonstrated that the FF policy is robust to the sensing failures and can hold the optimal motion. Attached video is also uploaded on youtube: https://youtu.be/zLL4uXIRmrE