Personal playground project to learn:
- JAX
- MuJoCo / MJX physics
- Brax PPO training loops
The goal is not a production-grade benchmark, but a fun sandbox for experimenting with custom rewards, weird poses, and humanoid behavior.
- Trains a Brax humanoid policy with PPO (JAX backend + MJX physics).
- Adds custom reward wrappers to push specific styles:
- Symmetry-aware walking
- "Naruto"-style arm pose while walking
- Renders rollouts to HTML for quick visual inspection.
- Includes extra MuJoCo assets (humanoid variants, shadow hand) for further experimentation.
assets/ MuJoCo XML assets (humanoids, scenes, shadow hand)
envs/
walking_env.py Custom Brax wrappers and reward shaping
scripts/
train_walking.py PPO training entry point (Brax)
evalaute.py Loads saved params and renders rollout to HTML
pose_extractor.py Utility to inspect joint poses in MuJoCo viewer
ppo_cleanrl.py Extra PPO experiment script (CleanRL-style)
models/ Saved model checkpoints (.pkl)
renders/ Rendered HTML rollouts
humanoid.ipynb Notebook playground