cagataydev/sac-unitree-g1-mujoco

SAC Unitree G1 — MuJoCo Locomotion Policy

A Soft Actor-Critic (SAC) policy trained for the Unitree G1 humanoid in MuJoCo simulation. Currently learning to balance — stays upright ~4 seconds and stumbles forward.

Trained entirely on a MacBook (CPU, no GPU, no Isaac Gym) using strands-robots.

Results

Metric	Value
Algorithm	SAC (Soft Actor-Critic)
Training steps	1.91M
Training time	~60 min (MacBook M-series, CPU)
Parallel envs	8
Network	MLP [256, 256]
Best reward	530
Mean distance	2.65m
Episode length	200/1,000 (4 seconds upright)
Status	Balancing + stumbling forward

Demo Video

Why It's Hard

The G1 has 29 DOF vs Go2's 12. Bipedal balance is fundamentally harder — the robot must coordinate hip, knee, ankle, and torso simultaneously while maintaining a tiny support polygon.

With more training (~5-10M steps, ~3 hours), it should learn to walk.

Usage

from stable_baselines3 import SAC

model = SAC.load("best/best_model")

obs, _ = env.reset()
for _ in range(1000):
    action, _ = model.predict(obs, deterministic=True)
    obs, reward, done, truncated, info = env.step(action)

Reward Function

reward = forward_vel × 5.0       # primary: move forward
       + alive_bonus × 1.0       # stay upright
       + upright_reward × 0.3    # orientation bonus
       - ctrl_cost × 0.001       # minimize energy
       - lateral_penalty × 0.3   # don't drift sideways
       - smoothness × 0.0001     # discourage jerky motion

Files

• best/best_model.zip — Best checkpoint
• checkpoints/ — All 100K-step checkpoints
• logs/evaluations.npz — Evaluation metrics
• g1_balancing.mp4 — Demo video

Environment

• Simulator: MuJoCo (via mujoco-python)
• Robot: Unitree G1 (29 DOF) from MuJoCo Menagerie
• Observation: joint positions, velocities, torso orientation, height (87-dim)
• Action: joint torques (29-dim, continuous)

License

Apache-2.0