Distributed control representation for manipulation tasks

摘要

Is there a robust basis for dexterous manipulation tasks? This approach relies on reusable control laws to put together manipulation strategies online. A demonstration is presented that suggests that the approach scales to the complexity of manipulation tasks. The compact control basis representation and the predictable behavior of the constituent controllers greatly enhances the construction of correct composition policies. This predictability allows reasoning about end to end problem solving behavior, which is not supported by methods employing less formal behavioral specifications. In those methods the designer must determine the composition policy, or the system must find it through random exploration. Our approach opens the composition problem to a large variety of control, planning, and machine learning methods. We are investigating formal methods that automatically generate composition policies from abstract task descriptions provided by the user. The generic character of the control basis not only improves generalization across task domains, but also appears to improve generalization across a variety of hardware platforms.