Figure has improved the control system of the Figure 02 humanoid robot to make its gait more human-like. The neural network-based system was fully trained in a simulation using reinforcement learning and then successfully transferred to real robots without additional tuning or retraining.

The training process took place in a physical simulator where the walking of a huge number of virtual copies of Figure 02 was simulated simultaneously. To increase the robustness and adaptability of the walking algorithm, the physical parameters of each virtual robot (mass, actuator characteristics, and so on) were varied in the simulation, and different operating conditions were simulated. The robots encountered different types of surfaces and external perturbations such as jolts or slips. To achieve a more human-like walking style, incentives were added to the reward system to encourage the robot to more closely mimic human movements, such as heel-to-toe foot placement and arm swings synchronized with leg movements. At the same time, the algorithm was trained to maintain a given walking speed, optimize energy consumption, and provide resilience to external influences