README.md

---
language:
- en
library_name: lerobot
pipeline_tag: robotics
tags:
- vision-language-action
- lerobot
inference: false
license: gemma
base_model: lerobot/pi05_base
---

# framepick base checkpoint (pi0.5 + subtask + frame re-expression)

Weights are bit-identical to [`lerobot/pi05_base`](https://huggingface.co/lerobot/pi05_base).
The checkpoint self-describes the **framepick** architecture (LeRobot plugin
`lerobot_policy_framepick`): `config.json` has `type: framepick`, and
`policy_preprocessor.json` carries the subtask tokenizer step plus the framepick
step with a NEUTRAL frame selector. Experiment knobs (which frames, selection
strategy) live in the training config, e.g.
`--policy.frame_selector.assignment='{"arm": "image2"}'`.

Generated by `lerobot_policy_pi05_subtask/scripts/prepare_base_checkpoint.py
--policy-type framepick`. See also `dgrachev/pi05_subtask_base` (same weights,
pi05_subtask architecture) and the `dgrachev/libero_extrinsics` dataset.