Monte-Carlo Tree Search for Efficient Visually Guided Rearrangement\n Planning

Abstract

We address the problem of visually guided rearrangement planning with many\nmovable objects, i.e., finding a sequence of actions to move a set of objects\nfrom an initial arrangement to a desired one, while relying on visual inputs\ncoming from an RGB camera. To do so, we introduce a complete pipeline relying\non two key contributions. First, we introduce an efficient and scalable\nrearrangement planning method, based on a Monte-Carlo Tree Search exploration\nstrategy. We demonstrate that because of its good trade-off between exploration\nand exploitation our method (i) scales well with the number of objects while\n(ii) finding solutions which require a smaller number of moves compared to the\nother state-of-the-art approaches. Note that on the contrary to many\napproaches, we do not require any buffer space to be available. Second, to\nprecisely localize movable objects in the scene, we develop an integrated\napproach for robust multi-object workspace state estimation from a single\nuncalibrated RGB camera using a deep neural network trained only with synthetic\ndata. We validate our multi-object visually guided manipulation pipeline with\nseveral experiments on a real UR-5 robotic arm by solving various rearrangement\nplanning instances, requiring only 60 ms to compute the plan to rearrange 25\nobjects. In addition, we show that our system is insensitive to camera\nmovements and can successfully recover from external perturbations.\nSupplementary video, source code and pre-trained models are available at\nhttps://ylabbe.github.io/rearrangement-planning.\n