Sim-to-Real Reinforcement Learning on Unitree G1: Domain Randomization & Actuator Net Calibration in Isaac Lab

Executive Conclusion: Learned Actuator Networks Close 80% of the Sim-to-Real Gap on Humanoids

RECIWO Labs' zero-shot deployment tests on the 23-DoF Unitree G1 prove that replacing ideal analytical PD torque equations in simulation with a 3-layer MLP Actuator Network trained on 15 minutes of real-world joint log data increases outdoor uneven-terrain walking survival from 38% to 96%. In NVIDIA Isaac Lab 2.2 on a single RTX 5090, 4,096 parallel humanoid environments converge to a robust 50Hz proprioceptive locomotion policy in just 18 minutes of wall-clock training.

1. Which Domain Randomization Parameters Actually Matter for Bipedal Sim-to-Real?

Over-randomizing every physics parameter creates an overly conservative "crouched" gait, while under-randomizing motor delay causes immediate ankle oscillation on real hardware.

2. Why Privileged Teacher-Student Distillation Beats Vision-Only Locomotion on Stairs

Even without turning on the head LiDAR, a blind proprioceptive student policy that encodes the last 300 milliseconds of joint position errors implicitly infers ground slope and stair step height the instant the toe contacts an obstacle. Fusing this 50Hz proprioceptive reflex loop with a 10Hz elevation map from the Livox Mid-360 LiDAR enables smooth 18cm stair climbing at 1.1 m/s.

[4,096 Parallel G1 Envs in Isaac Lab (RTX 5090)] ──(18 Mins PPO Training)──> [1.4MB ONNX Teacher-Student Policy]
                                                                                      │
                                                                                      ▼
[Unitree G1 Onboard Orin NX] ──> [50Hz Proprioceptive History (Last 15 Steps)] ──> [500Hz CAN-FD Joint PD Targets]