Control and Motion Planning for Compliant Humanoid Robots

本研究计划的目标是开发有能力、灵活和高效的类人机器人。尽管我和其他研究人员在过去十年中取得了巨大的进步，但我们离能够充当消防员或照顾老人的类人机器人还很遥远。其中一个挑战是机器人的运动和操作能力。在平坦平坦的地面上行走已经有了显著的进步。例如，2007年，hurouup马拉松项目的世界纪录是38分钟42米；2014年，世界纪录是45分钟跑240米。

The goal of this research program is the development of capable, flexible, and efficient humanoid robots. In spite of the great advances that other researchers and myself have achieved in the last decade, we are still far away from humanoid robots that can act as firemen or care for the elderly. One of the challenges are the locomotion and manipulation capabilities of the robots. Significant improvements have been made in walking on flat, even surfaces. For example, in 2007, the world record in the marathon event of HuroCup was 42 meters in 38 minutes; in 2014, the world record was 240 meter in 45 minutes. This proposal aims at investigating radically new approaches to the design, implementation, and control of humanoid robots. The majority of current humanoid robot designs are based on stiff structures: stiff links connected to gear boxes and electric motors. These designs work well for slow movements in structured environments, so that impacts between robots, environment, and humans can be predicted or entirely avoided. So nowadays humanoid robots can only walk or jog slowly. However, - similar to humans - walking over rough terrain, jumping, soft landing, climbing, and running robots must absorb large impacts and be able to store energy for short bursts at a later stage. In our research, we introduce some compliance into the design of the robots, so that are flexible. To this end, we replace stiff links with soft materials and add springs for short term energy storage. Changes in the design of the robot are only the first step. The new approach requires modifications to the low-level motor controllers, the active balancing and push recovery algorithms, and the motion planner. These systems cannot be looked at in isolation, since the low-level motor controllers determine what information the motion planner has available and the motion planner needs to send the correct control targets to the low-level motor controllers, all the while the robot is dynamically stable and balanced. We plan to automatically identify regions of control for the motor controllers that can be used as stepping stones by the motion controller and balancing algorithms. Recent advances in embedded processing power and better understanding of walking gaits and balancing reflexes have made this approach possible. The immediate outcome of this research will lead to faster locomotion, more dexterous manipulation, and safer robot designs. Long term, the coming robot revolution will affect almost everyone, even more than the PC and Internet revolution as computers will then be able to manipulate not just the virtual, but also the physical world. The research will result in the training of inter-disciplinary problem solvers that are able to develop sophisticated computer programs, electronics, and robot designs. These skills are crucially important in the new "Maker economy."