User Guide · Reference · CLI
CLI reference
Two binaries cover the everyday MetaFine loop: record.py for generating expert demos and core/policies/<policy>/evaluate.py for the three-dimensional diagnostic eval. The tables below summarise the canonical flags; for the live source of truth, every binary supports --help.
record.py
| Flag | Default | Notes |
|---|---|---|
--task | — | Atomic skill name (e.g. grasp_part). Mutually exclusive with --task-graph. |
--task-graph | — | Path to a task-graph YAML. |
--object | — | Asset id (required in single-skill mode). |
--trials | 1 | Number of successful episodes to write. |
--seed | 0 | Base seed; per-episode seeds are seed + idx. |
--max-steps | 400 | Per-episode simulator step budget. |
--camera | third_person | Override the recording camera rig. |
--render-mode | rgb_array | human opens the SAPIEN viewer. |
--output | demos/ | Override the output root. |
core/policies/<policy>/evaluate.py
Every policy ships the same evaluate CLI by convention; flags below are common across backbones.
| Flag | Default | Notes |
|---|---|---|
--task-graph | — | Path to a task-graph YAML. Required. |
--checkpoint | — | Path to the trained model weights. Required. |
--episodes | 30 | Episodes per cell of the eval grid. |
--seed | 0 | Eval seed (separate from training seed). |
--max-steps | 400 | Per-episode step budget. |
--camera | third_person | Camera rig (must match training). |
--dr-sweep | None | standard runs lighting + view + jitter; otherwise skip DR-AUSC. |
--smoothness | off | Record per-step joint state to compute jerk/vel-var/path-length. |
--render-mode | rgb_array | Set human for live visualisation. |
--output | ./ | Where results.json is written. |
Canonical source: python record.py --help and python core/policies/<policy>/evaluate.py --help. Flag names above match the canonical conventions; backbone-specific flags (LoRA rank, diffusion steps, chunk size) live in their own pages.