Neural Network Teaches Robots to Walk Like Humans

Researchers developed an end-to-end neural network trained with reinforcement learning, enabling robots to learn human-like walking in simulation and seamlessly transfer this capability to real hardware without additional tuning, achieving robust and natural locomotion across multiple robots.

• We introduced an end-to-end neural network trained with reinforcement learning for humanoid locomotion.
• Our robot learns to walk like a human via a high-fidelity physics simulator, and the trained policy transfers zero-shot to real hardware without additional tuning.
• By combining domain randomization in simulation and high-frequency torque feedback on the robot, our approach enables robust and human-like walking across the entire fleet of Figure 02 robots.