Blogs›Figure AI›Humanoid Locomotion Control

Humanoid Locomotion Control

posts

2025

Figure AI has developed an end-to-end neural network for humanoid locomotion, trained using reinforcement learning (RL) in a high-fidelity physics simulator. This system enables Figure 02 robots to walk naturally, mimicking human gaits with heel-strikes, toe-offs, and synchronized arm swings. The approach leverages domain randomization and kHz-rate torque feedback for robust sim-to-real transfer, allowing policies to generalize zero-shot to real hardware across the entire robot fleet without additional fine-tuning. This facilitates rapid engineering iteration and scalable commercial operations.

2025

Natural Humanoid Walk Using Reinforcement Learning

3/25/2025

Introduced an end-to-end neural network for humanoid locomotion trained with reinforcement learning (RL). The system utilizes a high-fidelity physics simulator with domain randomization and zero-shot sim-to-real transfer via kHz-rate torque feedback. The RL policy is optimized to mimic human walking reference trajectories, balancing gait style with velocity tracking, power consumption, and robustness to perturbations and terrain variations. This enables repeatable, human-like walking across the fleet of Figure 02 robots.