Commands and Scripts¶
This reference lists maintained public entrypoints. Run repository Python commands from the AM-Bench root after activating the Isaac Lab environment:
Commands that create an environment require Isaac Sim and a compatible NVIDIA GPU. Dataset-only validation and conversion do not launch Isaac Sim, but require their policy/data dependencies.
Environment commands¶
| Entrypoint | Purpose | Important arguments and behavior |
|---|---|---|
python scripts/environments/list_envs.py |
Print registered AM-Bench IDs, entrypoints, and config classes | Headless; launches Isaac Sim; no command-specific arguments |
python scripts/environments/zero_agent.py |
Continuously step an environment with neutral actions | required --task; optional --num_envs, --video, --camera_names, --disable_fabric, and AppLauncher options |
python scripts/environments/teleop_se3_agent.py |
Control an environment interactively | required --task; --num_envs 1, --teleop_device keyboard, and --sensitivity 0.2 defaults; device increments are integrated into absolute EE targets |
zero_agent.py and teleoperation run until the application closes. Use the bounded commands in Verify Installation and Run an Environment when a finite smoke test is required.
Recording and dataset commands¶
Script argument names in this section use their current underscore spelling.
| Entrypoint | Purpose | Required input | Primary output |
|---|---|---|---|
python scripts/data/record_demos_scripted.py |
Record task-aware scripted demonstrations | --task |
timestamped session containing env_cfg.yaml and canonical lerobot/ data |
python scripts/data/record_demos_teleop.py |
Record human teleoperation demonstrations | --task |
same canonical session layout |
python scripts/data/validate_lerobotdataset.py |
Statically validate finalized canonical data | --dataset_root; optional --repo_id and --target_hz |
console validation report and exit status |
python scripts/data/dp/lerobot_to_zarr.py |
Convert canonical data to UMI zarr | --input_path; optional --output_path |
.zarr.zip or zarr directory |
python scripts/data/dp/validate_zarr.py |
Validate a DP zarr conversion | positional zarr_path; optional --image_size |
console validation report and exit status |
python scripts/data/export_lerobot_to_openpi.py |
Export one or more canonical roots for the pinned OpenPI reader | --dataset_roots; optional --repo_id, --output_root, and prompt mapping |
derived local LeRobot v2.1 dataset |
Both recorders accept --dataset_root, --step_hz, --num_demos, --env_length_s, --camera_names, --video, --repo_id, --state_keys, and --task_prompt. The scripted recorder additionally accepts --num_envs, --inject_noise, and --save_failed_episodes; teleoperation accepts --teleop_device and --sensitivity.
Conversion outputs are derived artifacts. Do not overwrite the canonical source dataset. See Collect Demonstrations, Validate Datasets, and Dataset Format.
Learned-policy entrypoints¶
| Policy | Training or server entrypoint | Evaluation entrypoint |
|---|---|---|
| ACT | python -m ambench_learn.policies.act.train |
python -m ambench_learn.policies.act.eval |
| Diffusion Policy | python train.py from the pinned UMI directory with a Hydra config |
python -m ambench_learn.policies.dp.eval |
| OpenPI | pinned ext/openpi compute-stats, PyTorch training, and policy-server modules |
python -m ambench_learn.policies.pi.eval |
The three learned-policy evaluators use a shared kebab-case interface. All require --task; ACT and DP require --checkpoint, while OpenPI requires --prompt and a reachable server. See Configure Evaluation for exact shared defaults and the individual policy guides for training and policy-specific commands.
Completion evidence¶
Environment smoke tests must reach scene creation, reset, and stepping. Dataset validators must exit successfully and print their passing message. A learned-policy evaluation is complete only when eval_summary.json reports status: "completed" and the requested rollout count; see Evaluation Outputs.