Learning Dynamic Non-Prehensile Object Reorientation Via Reinforcement Learning

This work proposes a learning-based approach to dynamic non-prehensile object reorientation, enabling fast reorientation of large, grasp-infeasible objects using uni-manual manipulation. Our policy is trained in simulation via rein- forcement learning, utilizing a carefully designed observation space, action space, and reward function to reorient ran- domly sized cuboids with varied physical properties. Given an object model and a target rotation direction, the policy plans offline trajectories suitable for both simulation and real-world deployment. Although the policy is sensitive to modeling uncertainties, experimental results show that accurate modeling enables successful sim-to-real transfer across different objects and rotation directions. Videos are available at https: //tinyurl.com/DNPR2L.