Receding-horizon online optimization for dexterous object manipulation

We present a system for optimal control of dexterous manipulation. We use numerical optimization to automate the planning of the robot’s motion while reasoning about its effects on the object. The user specifies the high-level task (e.g., the desired effect on the object), and the optimization allows the robot to autonomously discover the low-level details of the motion that fulfils this task. Applying this optimization in real time allows the system to adapt the plan to changes in the environment and the task. We demonstrate the effectiveness of the integrated system using a four-finger robot platform manipulating a cylindrical object.