Sim-to-real policy transfer -- deploying policies trained in simulation in the real world -- is a promising paradigm for scaling robot manipulation without large-scale real-world data. However, transferring simulation-trained policies remains challenging due to discrepancies in c…
Exploiting the morphological symmetry of robotic systems, such as humanoid and quadruped robots, is a promising direction for improving robot learning. In deep reinforcement learning (DRL) for robot control, prior studies have leveraged such symmetry to improve learning efficienc…