Abstract:
Reinforcement learning has shown to be a powerful tool for decision-making problems. In this talk, we present the opportunities and challenges of enabling increasingly complex robot behavior with reinforcement learning. First, we present a system that combines reinforcement learning and extrinsic dexterity to solve a novel task of “occluded grasping”. To reach an occluded grasp pose of an object, the robot learns to interact with the object and the extrinsic environment to co-optimize pre-grasp and grasping motions. We demonstrate the generality of the learned policy across environment variations in simulation and evaluate it on a real robot with zero-shot sim2real transfer. Second, efficiently using offline data is important in reinforcement learning for robotics applications because data collection is expensive on real robots. We present a method for offline reinforcement learning which learns a Policy in the Latent Action Space (PLAS) to naturally avoid out-of-distribution actions. The method achieves competitive performance on continuous control benchmarks in simulation and a cloth sliding task on a physical robot.

Committee:
Prof. David Held (Chair)
Prof. Abhinav Gupta
Prof. Oliver Kroemer
Mohit Sharma