Troubleshooting & FAQ
failed to legalize operation 'mhlo.cholesky'
jax-metal (Mac GPU) is installed and JAX is using it. This physics does not run on Metal. Use a clean .venv installed from requirements-lock.txt; it must not contain jax-metal.
The first import takes very long or downloads something
MuJoCo Playground is trying to clone mujoco_menagerie. Run the empty-folder command from Setup.
ModuleNotFoundError: playground
Make sure the first argument to the scripts is the Open_Duck_Playground folder (external/Open_Duck_Playground). The scripts load the playground package from there.
Why python -I everywhere?
-I (isolated mode) stops Python from loading stray modules from the current folder or environment variables. It’s safer when you run inside a downloaded folder.
Training
Section titled “Training”Steps per second are very low
Try --devices 4, 6 and 8 with smoke_pmap.py. Heavy background work (video rendering, browser tabs) can halve the speed; in our run chunks ranged from 1,924 to 7,358 steps/s.
Nothing seems to happen after the first checkpoint
The first JAX compile takes a few minutes. After that each chunk takes ~9 minutes on our machine. checkpoints.csv and the steps= lines in the terminal show progress.
The duck hasn't walked for hours. Is something broken?
Probably not. Ours learned to stand in 11 minutes, then sat on a plateau for ~3 hours and only lifted its feet at 68.4 M steps. If episode return and the imitation term keep rising, it is learning.
Can I stop and resume training?
train_mac.py has no resume option. --restore <ckpt.pkl> in reward_hacking/train_hile.py continues from a checkpoint (policy and observation normalisation; value network from scratch).
The Mac went to sleep and training stopped
Closing the lid puts the Mac to sleep. Prefix long runs with caffeinate -i.
Results
Section titled “Results”My numbers differ slightly from yours
Evaluation runs in C MuJoCo and training in MJX; different JAX/MuJoCo versions or CPUs can cause small numerical differences. requirements-lock.txt and the pinned commit (b9be205) get you closest. Training itself can take a different path depending on seed and hardware.
Can I run this policy on a real Open Duck Mini?
The ONNX files have the upstream interface, but we have not tried hardware. The simulated motor limit is ~70 % above the servo’s catalog rating: part 3.
Is this the same as Microduck?
No. Microduck is a separate product by Pollen Robotics (part of Hugging Face); this repository uses Open Duck Mini v2. Differences and what you can do on a Mac: Microduck.