VLA Gallery / All models / Humanoid whole-body controllers

Figure RL 보행 제어기

Figure RL locomotion controller (Reinforcement Learning Walking)

Figure AI · 2025-03

VAHumanoid whole-body controllersNot disclosedRL policy

Simulates thousands of Figure 02 units with different physical parameters in parallel, learning years' worth of sim demonstrations in a few hours.

Paper

No paper

Backbone LLM: None
Vision encoder: —
Action head: kHz-level torque feedback control (RL policy)
Chunk: Not disclosed
Control rate: Not disclosed
Architecture (original summary): End-to-end neural network trained with RL (type not disclosed); kHz-level torque feedback control on the robot

Pretraining accelerators: Not disclosed (GPU-accelerated sim, name Not disclosed)
Pretraining duration: 'A few hours' (years' worth of simulated demonstrations)
Pretraining steps & batch: Not disclosed (parallel sim of thousands of Figure 02 units)
GPU/TPU-hours: Not disclosed
Fine-tuning: Not disclosed
Data type: Simulation·synthetic
Data scale: Human walking reference trajectories + parallel sim experience (scale not disclosed)
Main datasets: Not disclosed
Tasks & evaluation: 10 Figure 02 units walk with the same RL neural network
Robot: Figure 02

Reinforcement learning: Used (RL controller training) · End-to-end locomotion neural network trained with simulation RL
Base control: Humanoid legged (RL policy drives joints directly) · RL policy directly controls Figure 02 walking (trained with sim RL; output executed in a kHz torque closed loop)

Key features: Simulates thousands of Figure 02 units with different physical parameters in parallel, learning years' worth of sim demonstrations in a few hours. Transfers to hardware via domain randomization; 10 Figure 02 units walk with the same RL neural network.
Notes: 2025 locomotion controller separate from Helix. Simulator·network type·GPU not disclosed. Blog 2025-03-25. Not covered by verification file.
Verification: Not verified