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

FlagDefaultNotes
--taskAtomic skill name (e.g. grasp_part). Mutually exclusive with --task-graph.
--task-graphPath to a task-graph YAML.
--objectAsset id (required in single-skill mode).
--trials1Number of successful episodes to write.
--seed0Base seed; per-episode seeds are seed + idx.
--max-steps400Per-episode simulator step budget.
--camerathird_personOverride the recording camera rig.
--render-modergb_arrayhuman opens the SAPIEN viewer.
--outputdemos/Override the output root.

core/policies/<policy>/evaluate.py

Every policy ships the same evaluate CLI by convention; flags below are common across backbones.

FlagDefaultNotes
--task-graphPath to a task-graph YAML. Required.
--checkpointPath to the trained model weights. Required.
--episodes30Episodes per cell of the eval grid.
--seed0Eval seed (separate from training seed).
--max-steps400Per-episode step budget.
--camerathird_personCamera rig (must match training).
--dr-sweepNonestandard runs lighting + view + jitter; otherwise skip DR-AUSC.
--smoothnessoffRecord per-step joint state to compute jerk/vel-var/path-length.
--render-modergb_arraySet 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.