Superseded by grab_cup_2cam_v6. Same goal — removing the static pauses that make a trained policy freeze mid-task — but v5 cuts them by splitting 145 episodes into 498 shorter ones. That triples the action chunks that run off an episode end and are filled with a repeated last action, which openpi trains on as real: 2.4% -> 9.5% of the loss. v6 drops the same frames in place, leaving the episodes intact and that figure at 2.8%. Use v6.

grab_cup_2cam_v5 — SO-101, two cameras

LeRobot v3.0 dataset for the SO-101 arm: grasp a white cup, fixed and handeye cameras, 30 fps. This is grab_cup_2cam_v4 with the long standing-still stretches removed.

	v4 (source)	v5
episodes	146	498
frames	104,819	78,771 (−24.9%)
fps	30	30
median episode length	23.6 s	4.3 s
static frames	50.5%	33.8%
dead chunks	24.5%	0.0%

Task prompt (exact string): Grab the white cup

Why v5 exists

A policy trained on v4 drives the arm to a correct pose and then freezes there forever. The cause is in the data:

• 50.5% of v4 frames have the arm not moving; in 80% of those the gripper is commanded closed.
• A policy that sees one observation at a time has no clock, so "pause here for two seconds" and "stay here forever" are the same input — and the majority vote over the data says "stay".
• At rollout the loop closes on itself: the model emits the arm's own state as the action, the arm holds, the observation never changes again.

The failure unit is a dead chunk. pi0.5 predicts 16 actions at action_stride=2 — a window of 32 source frames. If the arm never moves in those 32 frames, the sample teaches "stay" and carries no information about when to resume. v4 is 24.5% dead chunks; v5 is 0.0%.

Why 498 episodes instead of 146

v5 has more episodes than v4, and they are much shorter. This is deliberate.

Interior frames could not simply be deleted. openpi requests action chunks with delta_timestamps, but LeRobot converts those into row-index offsets (get_delta_indices) clamped to the episode bounds — the lookup is by row, not by time. Deleting rows mid-episode would make a chunk span the gap and teach a jump the arm cannot make, and nothing would raise an error.

So each cut splits the episode in two. Every v5 episode is a contiguous window of one v4 episode; chunks stop at the boundary and are flagged action_is_pad, as at a real episode end. One v4 episode usually becomes 3–4 v5 episodes.

Two practical consequences:

• Videos are shared and were not re-encoded. Several episodes point at the same .mp4 with different videos/<key>/from_timestamp offsets. Do not assume one video per episode, and do not re-encode them.
• Retained short pauses sit at the start of the following segment, never at the end — a segment ending in stillness would itself produce dead chunks.

Pauses shorter than 32 frames were kept. A chunk starting inside one also contains the arm moving again, so "hold briefly, then go" stays representable.

Training with openpi

The repo ships the precomputed normalization statistics under openpi_assets/, so you do not have to recompute them.

1. Place the dataset where LeRobot looks for it. openpi calls LeRobotDataset(repo_id) without an explicit root, so it resolves under $HF_LEROBOT_HOME:

mkdir -p "$HF_LEROBOT_HOME/so101"
ln -s /path/to/grab_cup_2cam_v5 "$HF_LEROBOT_HOME/so101/grab_cup_2cam_v5"

The openpi config uses repo_id="so101/grab_cup_2cam_v5" — a local name, not this Hub id. Keeping it that way matters: asset_id is derived from repo_id and decides where the norm stats are looked up.

2. Drop the norm stats into the openpi checkout:

cp -r openpi_assets/so101 openpi/assets/pi05_so101_v5/
# -> openpi/assets/pi05_so101_v5/so101/grab_cup_2cam_v5/norm_stats.json

Do not reuse v4's norm stats. Removing half the standing-still frames shifts the state and action distributions. To regenerate instead: uv run python scripts/compute_norm_stats.py --config-name pi05_so101_v5.

3. Train with config pi05_so101_v5 (pi0.5, action_dim=32, action_horizon=16, 30k steps, batch 32, fsdp_devices=4):

uv run python scripts/train.py pi05_so101_v5 --exp-name <name>

4. Serve. --policy.config must match the config used for training, or serve_policy looks for norm stats under the wrong dataset name and fails with FileNotFoundError on norm_stats.json:

uv run scripts/serve_policy.py policy:checkpoint \
    --policy.config=pi05_so101_v5 --policy.dir=/path/to/checkpoint

Regenerating v5 from v4

python clean_dataset.py --src grab_cup_2cam_v4 --dst grab_cup_2cam_v5 \
    --max-static-frames 31 --write

--max-static-frames 31 must stay below the chunk window (action_horizon × action_stride = 32) or dead chunks reappear. At 31 it removes the least data that still drives them to zero.