XenderYang/microduck-rl-ydx-walk

MicroDuck RL — Velocity Walking (trained 4096 env x 6000 PPO)

Self-trained walking baseline for the open-source MicroDuck biped (Duck Together / RoboGo cloud project).

• Task: Mjlab-Velocity-Flat-MicroDuck (flat-ground velocity tracking)
• Training: mjlab + MuJoCo Warp + PPO, 4096 parallel envs x 6000 iterations, ~1h59m on NVIDIA RTX 5090
• Contract: ONNX [1, 61] -> [1, 14] (observation normalizer baked in via scripts/export.py)
• Result: mean reward ~81.5, mean episode length ~919.9, no NaN
• Note: this is an intermediate baseline (stable stand/walk, weaker straight-line speed tracking than the official alpha_walking). Full training config/log diff included under params/ and microduck_rl.diff.

Files

File	Description
model_5999.pt	final PyTorch checkpoint
final.onnx	exported ONNX policy [1,61] -> [1,14]
params/agent.yaml, params/env.yaml	training configuration
microduck_rl.diff	upstream repo diff (commit pin)

Upstream

• pollen-robotics/microduck
• pollen-robotics/microduck_rl (Apache-2.0)