Accelerating Grasp Exploration by Leveraging Learned Priors

摘要: The ability of robots to grasp novel objects has industry applications in e-commerce order fulfillment and home service. Data-driven grasping policies have achieved success learning general strategies for arbitrary objects. However, these approaches can fail which complex geometry or are significantly outside the training distribution. We present a Thompson sampling algorithm that learns given object with unknown using online experience. leverages learned priors from Dexterity Network robot planner guide exploration provide probabilistic estimates each stable pose object. find seeding policy Dex-Net prior allows it more efficiently robust grasps on Experiments suggest best attains an average total reward 64.5% higher than greedy baseline achieves within 5.7% oracle when evaluated over 300,000 runs across set 3000 poses.