Micro Duck RL — Cool Tricks Series 🦆
Training a series of cool tricks for the Micro Duck quadruped with reinforcement learning (mjlab + rsl_rl). This project is ongoing — more tricks will be trained and released here over time.
Tricks
🧱 Roulade Wall-Jump (released)
Full curriculum from 0.5cm → 5cm walls: approach → jump school (A3/A4) → 2cm wall school (BL1) → stand-up chain (BL2) → midroll → full-chain (B19) → steep dive (B20) → 5cm graduation (B24).
- 🎥 Demo:
demos/b24_5cm_best_run.mp4 - 🏋️ Checkpoint:
checkpoints/b24_5cm_fullchain/model_228468.pt(rsl_rl PPO actor-critic)
🔜 Coming soon
- Lean 9-stage curriculum checkpoints (training in progress)
- More cool tricks (backflip dives, stands, combos...)
Usage
git clone https://github.com/scchy/microduck-rl-cool-curriculum
# place checkpoint under logs/rsl_rl/... and resume / evaluate with mjlab_microduck CLI
Training scripts: scripts/train_curriculum_lean_to_5cm.sh in the GitHub repo.
Based on pollen-robotics/microduck_rl.