A model predictive approach for online mobile manipulation of non-holonomic objects using learned dynamics

Assistive robots designed for physical interaction with objects will play an important role in assisting mobility and fall prevention healthcare facilities. Autonomous mobile manipulation presents a hurdle prior to safely using real-life applications. In this article, we introduce framework based on model predictive control learned dynamics models of objects. We focus the specific problem manipulating legged such as those commonly found environments personal dwellings (e.g., walkers, tables, chairs). describe probabilistic method autonomous learning approximate these method, learn dynamic parameters small dataset consisting force motion data from interactions between robot object. Moreover, account multiple strategies by formulating planning mixed-integer convex optimization. The proposed considers hybrid system composed (i) choosing which leg grasp (ii) continuous applied forces manipulation. formalize our algorithm compensate modeling errors find optimal path manipulate object one configuration another. present results several various wheel configurations. Simulation experiments show that obtained are sufficiently accurate safe collision-free When combined algorithm, successfully moves desired pose while avoiding any collision.