Recently, Figure AI unveiled significant advancements in its humanoid robot locomotion, showcasing natural walking capabilities trained through reinforcement learning. This technology not only drastically improves the robot's movement speed but also marks a new milestone for AI-driven robotic control systems. The new generation robot, Figure02, boasts a walking speed of 2.68 miles per hour (approximately 1.2 meters/second), nearing the average human walking speed (around 3-4 miles per hour). This represents a nearly sevenfold increase compared to its predecessor, Figure01, which had a speed of 0.67 miles per hour.
Figure AI employs an innovative training method called "simulation-to-reality" (Sim-to-Real). By leveraging reinforcement learning in a high-fidelity simulated environment, engineers achieved the equivalent of years' worth of training data in just a few hours. After training, the neural network can be directly applied to the real robot without further adjustments, achieving what's known as "zero-shot" transfer. This method significantly shortens the development cycle while ensuring the technology's applicability in real-world settings.
Unlike traditional robotic control methods, Figure AI abandons rule-based heuristic design, instead relying entirely on an end-to-end neural network. This system, through autonomous learning, can adapt flexibly to complex tasks and environmental changes without requiring pre-programmed instructions. As a result, Figure02's gait is smoother and more natural, gradually approaching human walking style, although it hasn't yet reached a "perfect" human-like gait.
Furthermore, this neural network is deployed across Figure's entire robot fleet, with all robots operating under the same set of weight parameters, ensuring action coordination and technological consistency. This clustered application not only improves efficiency but also lays the foundation for future large-scale deployment.
Industry experts say Figure AI's progress demonstrates the immense potential of reinforcement learning in the field of humanoid robotics. The seamless transition from simulated training to real-world application accelerates technological iteration and makes it possible for robots to become more integrated into human life. In the future, with further gait optimization, Figure's robots are expected to take on more tasks in factories, homes, and even outdoor environments.