Towards human-like manipulation of robotic arms using reinforcement learning and stereo vision

Abstract

In recent years, a shift from classic fixed arm robotic manipulators towards mobile platforms with manipulators has been observed due to the increase in flexibility of tasks. More recently, the focus has shifted towards humanoid robotics. Whereas the dynamics of the legs and self stabilization of such bi-petal robots have been solved, the behavior of the humanoid arms needs further improvement to move more naturally and human-like. To overcome the hurdle of lack of natural arm movement, as a human does subconsciously, to keep balance while moving different parts of the body, we propose an experimental setup to collect arm and hand movement data from humans using a set of Color (RGB) and Color and Depth (RGBD) cameras via MediaPipe, but without needing any markers. A robotic arm manipulator with a custom soft robotics hand then mimics the movement and shows direct output for validation. In this way, several daily human manipulation tasks can be mimicked, recorded and used for training to eventually enable full autonomy in the future. We conducted experiments to quantify the accuracy of the mimicked pose, focusing on the end-effector pose (position and orientation) while monitoring the joint configuration. In our experiments, an accuracy of ±2.5cm was observed in all three axes (x,y,z); the orientation accuracy lies within ±10°. The demonstrator has been tested during several internal lab dry runs (for data collection) and public demos (showcasing robustness and longterm stability, e.g. during the Long night of research in Austria). Our work will serve as a basis for future multi-arm tracking and mimicking approaches.