pick3_raw8d_chunkrel

The pick3 episodes in raw 8D absolute-pose form, for training with chunk-relative actions. Same 554 episodes as SteveNguyen/grabette_pick3_graspproj_480 (can 166 / mustard 199 / cup 189), so the two are a controlled A/B on the action representation alone.

episodes / frames	554 / 91123
fps	50
camera	1 x 480x360
action	8D ['x', 'y', 'z', 'ax', 'ay', 'az', 'proximal', 'distal'] — ABSOLUTE pose (axis-angle) + gripper radians
observation.state	2D ['proximal', 'distal'] — gripper only

The action stats are RELATIVE, not absolute

meta/stats.json's action entry describes the chunk-relative distribution (chunk_size 50), not the absolute poses stored in the parquet. That is deliberate: the training pipeline converts to chunk-relative offsets before normalisation, and pi05 normalises ACTION with QUANTILES, so absolute stats would push the rotation channels outside [-1, 1]. The original absolute stats are kept under action_absolute.

So this dataset is only correct to train on with the chunk-relative processor step. Training on it as-is, with these stats, will mis-normalise.

How to use it

hf jobs uv run --flavor a100-large --timeout 24h -s HF_TOKEN \
    --env GRABETTE_CHUNK_RELATIVE=1 \
    --with "grabette-chunkrel @ git+https://github.com/pollen-robotics/grabette@<rev>#subdirectory=packages/grabette-chunkrel" \
    train.py -- --policy.type=pi05 --dataset.repo_id=SteveNguyen/pick3_raw8d_chunkrel ...

See docs/relative_actions_lerobot_native.md in the grabette monorepo for why the built-in --policy.use_relative_actions cannot be used here.

Provenance

Assembled from test_pick_can_200, test_pick_mustard_200 and grabette-dataset-cup-0107: clean (--keep_cameras cam0) -> resize 480x360 -> merge -> relabel tasks to the language strings -> relative action stats. No delta conversion: the delta representation is what this replaces. Jump repair spliced 25 translation and 45 rotation discontinuities across 2093 chunks.