Versatile Bipedal Locomotion and Walking-Running Transition: Coordinating Supervised Learning and Nonlinear Optimization

Online gait planning plays a crucial role for the locomotion of humanoid robots. While simplified models often fail to capture critical dynamic features of the robot’s motion, making online gait modifications with constrained nonlinear optimization in complex models is highly challenging with current computational power. This paper introduces a gait planning method that leverages supervised learning to expedite the gait planning and optimization process. Building upon our previous work, this paper extends a three-body model to the three-dimensional (3D) case. This model incorporates the angular momentum and height variation of body as well as the influence of leg motions, thus facilitating the generation of omnidirectional walking and running patterns. Due to the complexity of the model, an online gait planning modification is impractical. Therefore, supervised-learning is employed to train a policy derived from the model-based gait planning approach. This policy is then implemented online to produce versatile locomotion. Furthermore, the gradient of the trained neural network is utilized for nonlinear optimization of gait parameters, significantly improving the robot’s balance against external perturbations. The effectiveness of the proposed method is validated through a series of experiments conducted in simulation and on the real robot BHR-T, confirming its capability to generate adaptive walking and running motions in response to varying demands and disturbances. Note to Practitioners—The gait planning method presented in this paper addresses the practical challenge of real-time locomotion control in humanoid robots. This approach is beneficial for applications requiring quick adaptability to external disturbances and varying terrain. By integrating supervised learning with model-based nonlinear optimization, this method allows for the efficient generation and optimization of stable walking and running patterns. The trained neural network not only enhances the robot’s balance but also reduces the computational burden, making it feasible to implement in real-world scenarios. This method has been validated in both simulations and real-world tests, demonstrating its potential to improve the performance of humanoid robots in practical deployments.