Skip to content
open-duck-mac

Scripts and files

Every script prints its usage with --help.

Scripts

Section titled “Scripts”
Script What it does
duck, duck.py Command line: setup, doctor, try, eval, train, watch; calls the scripts below, standard library only (Try it in 5 minutes)
train_mac.py Brax PPO on the unchanged joystick.py env (flat terrain), BerkeleyHumanoidJoystickFlatTerrain hyperparameters, pmap over virtual CPU devices; logs, ONNX + params per checkpoint
onnx_export.py Brax policy params → ONNX with plain onnx (no TensorFlow/tf2onnx); also outputs loc_out/scale for noisy sampling
smoke.py, smoke_pmap.py Does the env compile and step, and how fast, on one device or D devices
eval_checkpoint.py Rollout of an ONNX policy in C MuJoCo with upstream MjInfer; distance, falls, foot steps; optional video
eval_grid.py The same with exploration noise for 9 seeds; how many fall, mean distance, optional 3×3 video
make_log.py Raw logs + rollout JSONs → train_log.csv, ckpt_summary.json, reward_curve.png
reward_hacking/hile_env.py Part 2’s reward file (VARIANTS)
reward_hacking/train_hile.py Training with reward variants, warm start with --restore
reward_hacking/make_hile_repo.py Adds massless floor-colliding boxes to a copy of Open_Duck_Playground
reward_hacking/render_ckpt_hile.py, render_grid_hile.py Part 2’s evaluators (training physics, extra measurements)
sim2real/gait.py From a measurement JSON: does the duck walk (foot oscillation, lift, left/right alternation)
sim2real/watch_ckpts.py Tries every new checkpoint of a training and logs whether it walks (--force, --once)
sim2real/eval_world.py eval_checkpoint.py + environment variables that change the world (FORCE, FORCE_RAND, FLOOR_FRIC, MASS_SCALE)

Result folders

Section titled “Result folders”
Folder Contents
results/2026-10-24/ Part 1: train_log.csv (one table), train_log_raw.csv, checkpoints.csv, ckpt_summary.json, train_metrics.jsonl (PPO losses, steps/s per chunk), run_config.json (PPO parameters), sim/*.json (every measurement), the 68.4 M and 72.4 M ONNX files, params/ckpt_068413440.pkl (for warm starts)
results/2026-10-24-reward-hacking/ Part 2: summary.json; per variant run.json, train_log.csv, train_metrics_ep.jsonl, onnx/, sim/

make_log.py rebuilds part 1’s train_log.csv and ckpt_summary.json from these files byte for byte.

Environment

Section titled “Environment”
  • Python 3.12, requirements-lock.txt (full environment), requirements.txt (direct dependencies: jax 0.6.2, mujoco / mujoco-mjx 3.3.3, brax 0.12.4, playground 0.0.5, onnx, onnxruntime, imageio).
  • Open_Duck_Playground commit b9be205ac64488c23504ca42e5ec790337adeec3.
  • Pretrained comparison: BEST_WALK_ONNX.onnx from Open_Duck_Mini commit b23317a.

Licence and credits

Section titled “Licence and credits”
  • Open Duck Mini v2 and Open_Duck_Playground by Antoine Pirrone and contributors: robot, MJCF model, joystick.py environment, rewards, reference motion (polynomial_coefficients.pkl), mujoco_infer.py and the pretrained BEST_WALK_ONNX. Open_Duck_Mini is Apache-2.0. Open_Duck_Playground’s source files carry Apache-2.0 headers, but its repository root had no LICENSE file at b9be205.
  • MuJoCo Playground, MuJoCo / MJX and Brax by Google DeepMind and contributors, Apache-2.0.
  • The code in this repository is released under the Apache License 2.0. results/ (logs, measurements, our ONNX policies) are released under the same licence.