Datasets:

lance-format
/

pusht-lerobot-lancedb

license: apache-2.0
task_categories:
  - robotics
language:
  - en
tags:
  - lerobot
  - pusht
  - diffusion-policy
  - robotics
  - lance
  - lerobot-lancedb
pretty_name: pusht-lerobot-lancedb (frames / JPEG)
size_categories:
  - 10K<n<100K

pusht-lerobot-lancedb (frames / JPEG)

A lerobot-lancedb-format Lance copy of lerobot/pusht. One row per frame; per-frame JPEG bytes stored inline.

Not to be confused with lance-format/lerobot-pusht-lance, which uses a different three-table schema (frames/episodes/videos) and is not loadable with lerobot-lancedb. This repo is specifically the layout produced by lerobot-convert-to-lance.

Load

from lerobot_lancedb import LeRobotLanceDataset

ds = LeRobotLanceDataset(repo_id="lance-format/pusht-lerobot-lancedb")
print(len(ds))                 # 25_650 frames
print(ds[0]["observation.image"].shape)  # torch.Size([3, 96, 96])

Lance streams per-frame bytes via byte-range fetches from the Hub — no full-dataset download.

Repro

pip install git+https://github.com/lancedb/lerobot-lancedb.git
lerobot-convert-to-lance \
    --repo-id=lerobot/pusht \
    --output=./pusht-lerobot-lancedb \
    --table-name=pusht-lerobot-lancedb \
    --jpeg-quality=95

Layout

pusht-lerobot-lancedb/
├── meta/                          # standard LeRobotDataset metadata sidecar
│   ├── info.json
│   ├── episodes.jsonl
│   ├── stats.json
│   └── tasks.jsonl
└── pusht-lerobot-lancedb.lance/   # single Lance table, one row per frame

Columns: episode_index, frame_index, index, timestamp, task_index, observation.state, action, observation.image (JPEG bytes, lossy q=95).

Size: ~60 MB. Companion video-blob layout (bit-exact pixels, ~8 MB) at lance-format/pusht-lerobot-lancedb-video.