Core Concepts
Humanoid-Gym is an open-source reinforcement learning framework designed to train locomotion skills for humanoid robots, enabling zero-shot transfer from simulation to the real-world environment.
Abstract
Humanoid-Gym is an open-source reinforcement learning (RL) framework based on Nvidia Isaac Gym, designed to train locomotion skills for humanoid robots. The key highlights of Humanoid-Gym include:
Zero-shot transfer from simulation to the real-world environment: The framework incorporates specialized rewards and domain randomization techniques to simplify the difficulty of sim-to-real transfer.
Sim-to-sim validation: Humanoid-Gym integrates a sim-to-sim framework from Isaac Gym to MuJoCo, allowing users to verify the trained policies in different physical simulations and ensure the robustness and generalization of the policies.
Verification on multiple humanoid robots: The framework has been successfully tested on RobotEra's XBot-S (1.2-meter tall) and XBot-L (1.65-meter tall) humanoid robots in a real-world environment with zero-shot sim-to-real transfer.
The workflow of Humanoid-Gym involves training agents using massively parallel deep reinforcement learning within Nvidia Isaac Gym, incorporating diverse terrains and dynamics randomization. The trained policies are then validated in the MuJoCo simulation environment, which is carefully calibrated to closely match the real-world dynamics. This comprehensive approach enables researchers to validate their training policies through sim-to-sim, significantly enhancing the potential for successful sim-to-real transfers.
Stats
The base poses of the robot are represented by six-dimensional vectors [x, y, z, α, β, γ], where x, y, z are the position coordinates, and α, β, γ are the orientation angles in Euler notation.
The joint position for each motor is represented by θ, and the corresponding joint velocity by ˙θ.
The policy network integrates proprioceptive sensor data, a periodic clock signal [sin(2πt/CT), cos(2πt/CT)], and velocity commands ˙Px,y,γ.
Quotes
"Humanoid-Gym features specialized rewards and domain randomization techniques for humanoid robots, simplifying the difficulty of sim-to-real transfer."
"Our open-source library features a sim-to-sim validation tool, enabling users to test their policies across diverse environmental dynamics rigorously."