Skip to content

Troubleshooting

Common failures and their fixes. Most are environment/version issues, not bugs in the package.

Installation and import

Symptom Cause Fix
AttributeError: ls_parallel was removed in MuJoCo Warp 3.9.1 at env build unpinned mujoco-warp (mjlab 1.2.0 sets an option removed upstream) install the pinned triple: mujoco==3.6.0 mujoco-warp==3.6.0 with mjlab==1.2.0 (see requirements)
ModuleNotFoundError: scipy mjlab 1.2.0 imports scipy without declaring it pip install scipy
pip install -e . cannot resolve safety_sb3 the pinned git tag is unreachable, or you are offline ensure network access to GitHub; to develop both repos, pip install -e the safety-stable-baselines checkout first (it satisfies the requirement)
Digit tasks warn-and-skip on import stock mjlab lacks the entity patch Digit needs to simulate expected — Digit imports fine; simulation needs the custom mjlab fork

GPU and rendering

Symptom Cause Fix
First run hangs for ~1 min with no output Warp is JIT-compiling kernels for your GPU arch wait — it is a one-time, per-machine compile; the cache is reused afterward
gladLoadGL error / OpenGL context error at the first video no display and no offscreen GL context export MUJOCO_GL=egl (headless NVIDIA EGL) before running
wandb.Video requires moviepy video encoder missing pip install moviepy imageio
Out-of-memory at env build --num-envs too high for your GPU lower --num-envs; Go2 tasks want ~4–9 GB at 2–3k envs

Checkpoints and evaluation

Symptom Cause Fix
ModuleNotFoundError / AttributeError unpickling a .zip checkpoint trained on an older version (learner class renamed) retrain on v0.4.0, or pin both repos to the old tag (see release notes)
argparse error on an eval command from older docs retired flags (--twin, --nominal, --nominal-from-twin) use the v0.4.0 names: --safety-policy, --task-policy, --safety-only (see evaluation)
Repeat eval runs give different numbers evaluation is not bit-reproducible (obs noise + non-deterministic Warp) compare configurations over several runs, not a single run
A run looks "converged" but performs poorly a stalled curriculum reads like converged training in the reward curve watch the env/Curriculum/* logger keys, not just ep_rew_mean

Still stuck?

  • Re-check the compatibility matrix — most issues are a version mismatch in the sim stack.
  • Re-run the installation verification commands (CPU import, then the GPU training smoke test) to localize the failure.