User Guide · Policies · π0 / π0.5

π0 & π0.5

MetaFine supports both π0 and its successor π0.5 — flow-matching action experts on top of PaliGemma vision-language backbones. Both consume the LeRobot export.

What they are

π0 and π0.5 — flow-matching VLA policies. Vendored under core/policies/pi0 and core/policies/pi05. π0.5 closed-loop inference in the MetaFine simulator is verified.

Data

Both consume a LeRobot dataset. Run record → merge → replay → convert_to_lerobot (see Quickstart).

Install

$ pip install -e core/policies/pi05

Evaluate (π0.5, verified)

The real evaluate.py is a standalone argparse script — no --task-graph adapter. There are two modes (general-task vs specific object/part); the full flag reference is core/policies/pi05/README.md:

$ python core/policies/pi05/evaluate.py \
    --policy-path /path/to/pretrained_model \
    --env-id grasp_part \
    --object-name 100221 \
    --part-name cap \
    --obs-mode rgb \
    --control-mode pd_joint_delta_pos \
    --n-episodes 50 \
    --device cuda \
    --task "Grasp the cap of the bottle." \
    --record-dir eval_out --save-video

For the staged diagnostic (semantic-intervention / object-swap — the Understanding protocol):

$ bash core/policies/pi05/run_eval_three_stage.sh --policy-path ... --env-id ... --task "..."

π0 follows the same shape; see core/policies/pi0/README.md. Canonical source: python core/policies/pi05/evaluate.py --help.