When using a tool, the grasps used for picking it up, reposing, and holding it in a suitable pose for the desired task could be distinct. Therefore, a key challenge for autonomous in-hand tool manipulation is finding a sequence of grasps that facilitates every step of the tool use process while continuously maintaining force closure and stability. Due to the complexity of modeling the contact dynamics, reinforcement learning (RL) techniques can provide a solution in this continuous space subject to highly parameterized physical models. However, these techniques impose a trade-off in adaptability and data efficiency. At test time the tool properties, desired trajectory, and desired application forces could differ substantially from training scenarios. Adapting to this necessitates more data or computationally expensive online policy updates. In this work, we apply the principles of discrete dynamic programming (DP) to augment RL performance with domain knowledge. Specifically, we first design a computationally simple approximation of our environment. We then demonstrate in physical simulation that performing tree searches (i.e., lookaheads) and policy rollouts with this approximation can improve an RL-derived grasp sequence policy with minimal additional online computation. Additionally, we show that pretraining a deep RL network with the DP-derived solution to the discretized problem can speed up policy training.
This paper presents a hierarchical framework for planning and control of in-hand manipulation of a rigid object involving grasp changes using fully-actuated multifingered robotic hands. While the framework can be applied to the general dexterous manipulation, we focus on a more complex definition of in-hand manipulation, where at the goal pose the hand has to reach a grasp suitable for using the object as a tool. The high level planner determines the object trajectory as well as the grasp changes, i.e. adding, removing, or sliding fingers, to be executed by the low-level controller. While the grasp sequence is planned online by a learning-based policy to adapt to variations, the trajectory planner and the low-level controller for object tracking and contact force control are exclusively model-based to robustly realize the plan. By infusing the knowledge about the physics of the problem and the low-level controller into the grasp planner, it learns to successfully generate grasps similar to those generated by model-based optimization approaches, obviating the high computation cost of online running of such methods to account for variations. By performing experiments in physics simulation for realistic tool use scenarios, we show the success of our method on different tool-use tasks and dexterous hand models. Additionally, we show that this hybrid method offers more robustness to trajectory and task variations compared to a model-based method.