Efficient Learning of Goal-Oriented Push-Grasping Synergy in Clutter

Interleaving Monte Carlo Tree Search and Self-Supervised Learning for Object Retrieval in Clutter

“…A human in-the-loop solution was proposed in [34] help with searching for objects in clutter. A deep Q-Learning method [35] considers a similar task and setup but uses additional primitives such as sliding objects from the top. Our work partially builds on [36], which explores the use of MCTS for the same object retrieval problem.…”

“…The results demonstrate that the proposed method significantly outperforms MCTS [36] in terms of time efficiency while returning plans of equal quality. The plans returned by the proposed technique contain fewer actions and yield higher success rates than those returned by the purely learningbased solution presented in [35]. Training and evaluation are completed on a machine with an Intel i7-9700K CPU and an Nvidia GeForce RTX 2080 Ti.…”

“…In simulation, we benchmark on 22 adversarial test cases from [36] (Fig. 5) and 10 from [24], [35]. Here "adversarial" means that at least one push action has to be executed for a grasp action to be feasible (insert gripper without collision).…”

“…Baseline Methods. We compare with three methods: 1) A self-supervised reinforcement learning method denoted as go-PGN [35], which trains a grasp DQN and a push DQN then selects an action with the highest Q-value out of the two networks to execute. 2) MCTS as described in Section IV-B.…”

“…Simulation Studies. We ran our method and the three alternative methods on 22 cases [36] and 10 cases [24], [35], in simulation first. Tables I and II show the overall performance of the four methods, where MCTS based methods are limited to a budget of 50 iterations per test case.…”