FAQ#
Installation#
No module named 'cv2' (or scipy, mink)#
The environment was synced without the extra that provides the module. Name
every extra in one uv sync, as in Installation.
Rendering#
Rendering fails on a headless Linux machine#
MuJoCo defaults to a windowed OpenGL context, which needs a display. Render through EGL instead:
export MUJOCO_GL=egl
Rendering fails on macOS with MUJOCO_GL=egl#
macOS has no EGL. Leave MUJOCO_GL unset and drop it from the commands in
these pages.
Environment and demonstrations#
Can I set reset_warmup_steps to 0 to save time?#
No. The demonstrations start after the 200-step warmup, so a run without it does not match them and is not official.
Which episode length should I use?#
Use the length make(task) sets (TASKS[task].episode_length). The
episode_length in a batch’s metadata.json is the collector’s cap
(Episode budgets).
get_demos() says the demonstrations do not match this env#
The loader compares the dataset’s robot, backend, downsample rate, cameras and
observation and action dimensions with the env. A camera override such as
camera_keys=("head",) is the usual cause. Build the env with
make(task) or make_gym(task), or re-render the dataset for another camera
setup with bigym-rerender-lerobot.
How do I work offline?#
Download once, then tell the Hub client to stay on its cache:
uv run bigym-download --all
export HF_HUB_OFFLINE=1
Evaluation#
My result has no leaderboard record#
The runner writes a record only for an official run with 100 episodes.
result["protocol_violations"] lists every result-affecting override.
Official overrides lists the
ones that keep a run official.
Are my numbers comparable with someone else’s?#
Official results are comparable when their substrate_version matches. See
Determinism and versioning.
VR collection#
Can I collect demonstrations on macOS?#
No. The collector needs an OpenXR runtime, and macOS has none. Use Linux with WiVRn and a Quest 3, as in Collecting demonstrations.
Coding-agent benchmark#
bigym-agent run refuses my model#
Each harness runs only its own vendor’s models (codex for OpenAI, claude
for Anthropic). Switch the harness to match the model.
My Claude Code subscription session stops authenticating mid-run#
A copied host .credentials.json shares one short-lived token between host
and container, and one refresh revokes the other. Create a dedicated token for
container sessions:
claude setup-token # then set CLAUDE_CODE_OAUTH_TOKEN