Sampling-Based Planning and Local Reactive Strategies for Environment Manipulation

We address the motion planning problem for a single robot that manipulates passive objects in an otherwise static environment. We handle uncertainties in passive object responses while abstracting their dynamics by creating a roadmap that encodes the possible passive object responses. The robot extracts reference paths from the passive roadmap to plan approaching motions with an active roadmap. When the robot acts on the object, it receives response feedback and reacts accordingly. We apply this approach to pushing, a basic manipulation primitive. We propose four different pushing strategies, one of them using reinforcement learning. Our simulations show the effectiveness of the approach and of each pushing strategy.