Reinforcement Learning

Reinforcement Learning(RL)

Robot reinforcement learning enables robots to learn control policies through trial-and-error interactions with their environment, making it effective for solving complex control problems in a data-driven way. Our lab studies robot reinforcement learning with a focus on integrated locomotion and manipulation. In particular, we build a mobile manipulator platform based on Unitree Go2 and WidowX, redefine the MDP to reflect the changed dynamics, and train new RL policies for coordinated control. We also develop an autonomous drone soccer environment in Isaac Lab to study flight control and cooperative multi-agent behaviors using reinforcement learning. In addition, we explore reinforcement learning for autonomous lunar rover navigation and terrain-adaptive driving in 3D simulation environments.

Overall, our research aims to improve autonomy, adaptability, and real-world applicability across diverse robotic platforms through simulation-based reinforcement learning.

This research focuses on the reinforcement learning-based integrated control of locomotion and manipulation for quadruped robots, rather than treating them as separate tasks. In particular, an expert policy trained in simulation is transferred to the real robot through DAgger and policy distillation, enabling stable navigation and manipulation in complex environments. The system further employs 5D command-based RL control to coordinate both the robot body and arm, improving its ability to perform long-horizon tasks that require continuous movement and object interaction.

Overall, this work demonstrates the potential of Sim-to-Real reinforcement learning for unified locomotion-manipulation control in quadruped robotic platforms.

Google Sites

Report abuse