Experience Guided Achievable Push Plan Generation for Passive Mobile Objects

I. OVERVIEW We contribute an algorithm that utilizes the observed effects of previously experimented and memorized pushing moves to generate achievable push plans for passive mobile objects (http://youtu.be/V2gf39mYepI). The most common scenarios in the push planning literature are pushing of a disk or box shaped quasi-static object with a smaller disk shaped or even point-sized robot, and the mathematical models are based on the single contact point of these two primitiveshaped entities. Our approach differs in the sense that • we deal with complex 3D real world objects that may contact the robot on more than a single point (Figure 1(c)), • the manipulated objects continue moving freely for a while after the push is ceased, • manipulation is performed in an environment crowded with obstacles (Figure 1(a)), • no explicit physical or mathematical model is constructed or learning-based mapping is built; only the pushing motions performed in the past and their corresponding observed effects are used for planning and execution (Figure 1(d)), • the push plan is generated by expanding an RRT tree based on achievable and non-colliding object trajectories (Figure 1(b)); hence, the probability of success is increased.