Motion Planning and Control of Dynamic Humanoid Locomotion

Inspired by human, humanoid robots has the potential to become a general-purpose platform that lives along with human. Due to the technological advances in many field, such as actuation, sensing, control and intelligence, it finally enables humanoid robots to possess human comparable capabilities. However, humanoid locomotion is still a challenging research field. The large number of degree of freedom structure makes the system difficult to coordinate online. The presence of various contact constraints and the hybrid nature of locomotion tasks make the planning a harder problem to solve. Template model anchoring approach has been adopted to bridge the gap between simple model behavior and the whole-body motion of humanoid robot. Control policies are first developed for simple template models like Linear Inverted Pendulum Model (LIPM) or Spring Loaded Inverted Pendulum(SLIP), the result controlled behaviors are then been mapped to the whole-body motion of humanoid robot through optimization-based task-space control strategies. Whole-body humanoid control framework has been verified on various contact situations such as unknown uneven terrain, multi-contact scenarios and moving platform and shows its generality and versatility. For walking motion, existing Model Predictive Control approach based on LIPM has been extended to enable the robot to walk without any reference foot placement anchoring. It is kind of discrete version of “walking without thinking”. As a result, the robot could achieve versatile locomotion modes such as automatic foot placement with single reference velocity command, reactive stepping under large external disturbances, guided walking with small constant external pushing forces, robust walking on unknown uneven terrain, reactive stepping in place when blocked by external barrier. As an extension of this proposed framework, also to increase the push recovery capability of the humanoid robot, two new configurations have been proposed to enable the robot to perform cross-step motions. For more dynamic hopping and running motion, SLIP model has been chosen as the template model. Different from traditional model-based analytical approach, a data-driven approach has been proposed to encode the dynamics of the this model. A deep neural network is trained offline with a large amount of simulation data based on the SLIP model to learn its dynamics. The trained network is applied online to generate reference foot placements for the humanoid robot. Simulations have been performed to evaluate the effectiveness of the proposed approach in generating bio-inspired and robust running motions. The method proposed based on 2D SLIP model can be generalized to 3D SLIP model and the extension has been briefly mentioned at the end.