Closed-loop RGB-D SLAM multi-contact control for humanoid robots

Abstract

We report the integration of a state-of-the-art dense 6D simultaneous localisation and mapping (D6DSLAM) to close the QP control loop on an humanoid robot in multi-contact motions. Our multi-contact planning defines desired contacts based on 3D (CAD) models, and generates a reaching plan. Registration of the 3D model onto an RGB-D key-frame graph representation of the explored environment allows using visual odometry to make real-time adjustments to the reaching plan, leading to improved in robustness from a wide range of perturbations. Extensive results are presented on various complex tasks using the HRP-2Kai humanoid robot including valve and car's steering-wheel grasping, multi-contact stair climbing from approximate initial humanoid poses.