Sliding-Window Planning for Realtime Motion Optimization in Humanoid Robots

Abstract

Autonomous robotic locomotion, particularly in the context of humanoid robots, has been an ongoing research topic for several decades. In this context, the model-based three-dimensional Divergent Component of Motion (3D-DCM) motion planning framework demonstrates comprehensive capabilities in planning dynamically consistent center-of-mass trajectories for agile motions. However, current methods for 3D-DCM motion planning either assume a constant Centroidal Angular Momentum (CAM) throughout the planned motion, are only able to roughly estimate its influence, or are not applicable to long, incrementally built motion plans as required, e.g., for shared-autonomy walking. This thesis addresses this issue by presenting a sliding-window approach for CAM -based optimization of the centroidal dynamics within the 3D-DCM framework. The multi-body dynamics of the robot are thereby simulated for short segments of the motion plan prior to their actual execution. These simulated preview windows enable a real-time capable motion plan optimization, even in shared autonomy scenarios, where short-term plan extensions are common and input time delays are undesirable. Furthermore, a conceptual path-search algorithm is proposed that is envisioned to enable short-term replanning of stepping sequences, for example, to avoid obstacles given a set of feasible alternative contact areas. The CAM -based optimization is validated both in simulation and reality on the Torque-Controlled Humanoid Robot (TORO) of the German Aerospace Center (DLR).