Learning Hybrid Object Kinematics for Efficient Hierarchical Planning Under Uncertainty

摘要: Sudden changes in the dynamics of robotic tasks, such as contact with an object or latching a door, are often viewed inconvenient discontinuities that make manipulation difficult. However, when these transitions well-understood, they can be leveraged to reduce uncertainty aid manipulation—for example, wiggling screw determine if it is fully inserted not. Current model-free reinforcement learning approaches require large amounts data learn leverage dynamics, scale poorly problem complexity grows, and do not transfer well significantly different problems. By contrast, hierarchical POMDP planning-based methods via plan decomposition, work on novel problems, directly consider uncertainty, but rely precise hand-specified models task decompositions. To combine advantages opposing paradigms, we propose new method, MICAH, which given unsegmented object’s motion under applied actions, (1) detects changepoints model using action-conditional inference, (2) estimates individual local their parameters, (3) converts them into hybrid automaton compatible planning. We show MICAH more accurate robust noise than prior approaches. Further, planner demonstrate learned rich enough used for performing tasks objects ways encountered during training.