Reference
A one-page map of the package: where each piece lives, its main entry point, and
the section that documents it.
Module map
| Package |
Role |
Docs |
Main entry point |
task_generation/ |
generate frozen task scenes |
Task Generation |
python -m maniguard.task_generation.<name>_pipeline · …run_benchmark |
data/teleop/ |
SO-101 / GELLO human teleop |
Data Collection |
python -m maniguard.data.teleop.{so101,gello}_franka_teleop |
data/datagen/ |
scripted sim demo collection → LeRobot |
Scripted datagen |
python -m maniguard.data.datagen.{driver,sweep,to_lerobot} |
data/playback.py |
render teleop HDF5 → SFT obs |
Data Collection |
python -m maniguard.data.playback |
data/lerobot/ |
sim teleop HDF5 → LeRobot multitask export |
Sim teleop → LeRobot |
python -m maniguard.data.lerobot.multitask_lerobot_export |
data/real_teleop/ |
real npz → LeRobot (DROID joint) |
Real teleop |
python -m maniguard.data.real_teleop.real_teleop_to_droid |
data/scene/ |
benchmark-set prep |
Evaluation |
python -m maniguard.data.scene.{benchmark_repair,trim_scene_to_room,rewrite_scene_robot} |
eval/ |
policy benchmark loop |
Run the benchmark |
bash scripts/eval_family.sh <family> · python -m maniguard.eval.benchmark --config <yaml> |
serve/ |
websocket policy servers (one per model family) |
Run the benchmark |
python -m maniguard.serve.{openpi,gr00t,smolvla}_native |
utils/ |
LTL safety, task_spec, geometry |
LTL safety |
(library) |
envs/ |
scene registry + frozen-snapshot runtime |
Environment layer |
(library) |
object_states/ |
Dropped, Upright |
LTL safety |
(library) |
_omnigibson_patches.py |
runtime OmniGibson patches |
OmniGibson patches |
applied on import maniguard |
Environment variables
| Variable |
Effect |
MANIGUARD_SKIP_OMNIGIBSON_PATCH=1 |
skip all runtime patches (no simulator needed) |
MANIGUARD_SKIP_LONGFINGER=1 |
skip the long-finger Franka asset patch |
MANIGUARD_AG_SUBSTEP_INTERVAL=N |
fire assisted-grasp once per N physics substeps |
MANIGUARD_RUNTIME_PYTHON |
torch-capable interpreter for frozen-task replay |
OMNIGIBSON_HEADLESS=1 · VK_ICD_FILENAMES |
headless GPU rendering |
OMNIGIBSON_DATA_PATH |
override the BEHAVIOR dataset root |
CUDA_VISIBLE_DEVICES |
GPU selection |
Running tests
conda activate behavior
pytest tests/ -v # maniguard-side LTL + task-gen + eval unit tests
See also