機器人運動員

The goal of this project is to create novel algorithms and technologies to support high accuracy mobile manipulation tasks, i.e., tasks that require an accuracy of at least 0.5mm. The proposed research uses robot athletes, in particular robot archery, as benchmark applications. Archery is an interesting and challenging benchmark problem. The robot must pinpoint the target, must pick up the bow and arrow and then nook the arrow. Then the robot must synchronize the motions of its torso, bow and shooting arm to generate maximum force for drawing the bow. Draw weights for olympic archers are about 20kg. Finally, the robot must compensate for external factors such as wind and temperature and release the arrow smoothly. The perception and motion components both require high accuracy under load and popular approaches in the research literature fail at this task. For perception, I propose a histogram of oriented gradients (HoG) based approach to find the coarse location of the target. The position of the target is then improved using local features to find the exact center of the target and to map it from camera to 3D world coordinates. I extend Voxnet and V-CNN1 deep learning architectures for object (e.g., the bow and arrow) recognition to improve their accuracy. Using multimodal sensor input from video camera, lidar scanner, and RGBD cameras, the proposed algorithm will provide accurate estimates of the bow, arrow, and drawstring. Common approaches to the inverse kinematics and dynamics of the robot lead to inaccuracies when the robot moves under heavy load (e.g., when trying to aimdraw the bow for long distance shots). I augment our inverse kinematics engine with a deep learning network to finetune the IK engine under load. The motion planner must be able to manipulate the bow and arrow so that when released, the arrow has a clean path to the target. This plan must be adapted based on the relative orientation and distance between the robot and the target.The robot must draw the bow under high load and keep it aligned with the target.Lastly, the robot must compensate for external factors in the environment such as wind and temperature. Since many parameters (e.g., backlash in the gears, stiffness of the bow, wind),are unknown, I propose a reinforcement learning based approach. My students and I have previously demonstrated the usefulness of reinforcement learning in learning motion plans to manipulate simple planar objects. However, standard DRL approaches are difficult to apply in complex multi-dimensional scenarios. In this research, I propose an extension to the soft actor critic approach. We will demonstrate the effectiveness of our research during international robot competitions as well as several field trials targeted at Taiwan sports and industry.The goal of the robot archery application is to develop a robot archer that is able to compete against Olympic caliber archers. This represents a significant step in robot development since it will be the first time that robots can beat humans under official competition rules.