lance-format/lerobot-pusht-lance

LeRobot PushT (Lance Format)

A Lance-formatted version of lerobot/pusht — the canonical PushT benchmark from the Diffusion Policy paper — packaged using the same three-table layout as lance-format/lerobot-xvla-soft-fold so consumers can flip between datasets without changing code. Available directly from the Hub at hf://datasets/lance-format/lerobot-pusht-lance/data.

Key features

• Three-table layout — frames, episodes, videos — so frame-level training, episode-level trajectory work, and raw video access live side-by-side without scattered parquet shards or sidecar MP4 directories.
• Inline MP4 segments in episodes.lance (one blob per camera, with from_timestamp / to_timestamp bounds) and full source MP4s in videos.lance, all surfaced as lazy BlobFile handles via take_blobs so metadata scans never read the bytes.
• Frame-level observations and actions in frames.lance with stable episode_index, frame_index, and index columns for joining or temporal iteration.
• Schema-evolution friendly — add alternate camera streams, language annotations, or model predictions later without rewriting the data.

Tables

Table	Rows ~	Purpose
frames.lance	one row per frame	Per-frame observations, actions, episode/task indices
episodes.lance	one row per episode	Full per-episode trajectories plus per-camera MP4 segment blobs and timestamp bounds
videos.lance	one row per source MP4	Raw source video blobs and file-level provenance (path, size, sha256)

Use frames.lance for low-level training (loss-per-timestep, state-conditioned policies). Use episodes.lance when you need the full trajectory and the matching video segments together. Use videos.lance when you want direct access to the original encoded video files.

Schemas

frames.lance

Column	Type	Notes
observation_state	list<float32>	Robot state vector for that frame
action	list<float32>	Action vector for that frame
timestamp	float	Canonical frame timestamp (seconds)
frame_index	int64	Frame index within episode
episode_index	int64	Parent episode id
index	int64	Global frame index
task_index	int64	Task id

episodes.lance

Column	Type	Notes
episode_index	int64	Episode id
task_index	int64	Task id
fps	int32	Frame rate of the episode video segments
timestamps	list<float32>	Per-frame timestamps
actions	list<list<float32>>	Per-frame action vectors
observation_state	list<list<float32>>	Per-frame robot state vectors
<camera>_video_blob	large_binary (blob-encoded)	Inline MP4 segment for each camera, read lazily via take_blobs
<camera>_from_timestamp	float64	Segment start time
<camera>_to_timestamp	float64	Segment end time

videos.lance

Column	Type	Notes
camera_angle	string	Camera key
chunk_index, file_index	int32	IDs parsed from the source path
relative_path, filename	string	Provenance
file_size_bytes	int64	Source MP4 size
sha256	string	SHA256 of the MP4 bytes
video_blob	large_binary (blob-encoded)	Raw source MP4 bytes

Pre-built indices

None bundled. Build indices on a local copy if a workload calls for them — e.g., a BTREE on frames.episode_index for fast episode lookup, or a vector index after attaching observation embeddings via Evolve.

Why Lance?

1. Blazing Fast Random Access: Optimized for fetching scattered rows, making it ideal for random sampling, real-time ML serving, and interactive applications without performance degradation.
2. Native Multimodal Support: Store text, embeddings, and other data types together in a single file. Large binary objects are loaded lazily, and vectors are optimized for fast similarity search.
3. Native Index Support: Lance comes with fast, on-disk, scalable vector and FTS indexes that sit right alongside the dataset on the Hub, so you can share not only your data but also your embeddings and indexes without your users needing to recompute them.
4. Efficient Data Evolution: Add new columns and backfill data without rewriting the entire dataset. This is perfect for evolving ML features, adding new embeddings, or introducing moderation tags over time.
5. Versatile Querying: Supports combining vector similarity search, full-text search, and SQL-style filtering in a single query, accelerated by on-disk indexes.
6. Data Versioning: Every mutation commits a new version; previous versions remain intact on disk. Tags pin a snapshot by name, so retrieval systems and training runs can reproduce against an exact slice of history.

Load with datasets.load_dataset

You can load Lance datasets via the standard HuggingFace datasets interface, suitable when your pipeline already speaks Dataset / IterableDataset or you want a quick streaming sample. Each Lance table is a separate datasets config.

import datasets

hf_ds = datasets.load_dataset("lance-format/lerobot-pusht-lance", split="frames", streaming=True)
for row in hf_ds.take(3):
    print(row["episode_index"], row["frame_index"], row["action"])

Load with LanceDB

LanceDB is the embedded retrieval library built on top of the Lance format (docs), and is the interface most users interact with. Each .lance file in data/ is a table — open by name. The same handles are used by the Search, Curate, Evolve, Train, Versioning, and Materialize-a-subset sections below.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/lerobot-pusht-lance/data")

frames    = db.open_table("frames")
episodes  = db.open_table("episodes")
videos    = db.open_table("videos")
print(len(frames), len(episodes), len(videos))

Load with Lance

pylance is the Python binding for the Lance format and works directly with the format's lower-level APIs. Reach for it when you want to inspect dataset internals — schema, scanner, fragments, the list of pre-built indices — or when you need the blob-level take_blobs entry point that streams MP4 bytes lazily from inline storage.

import lance

ds = lance.dataset("hf://datasets/lance-format/lerobot-pusht-lance/data/frames.lance")
print(ds.count_rows(), ds.schema.names)
print(ds.list_indices())

Tip — for production use, download locally first. Streaming from the Hub works for exploration, but heavy random access to video segments and any kind of indexed search are dramatically faster against a local copy:

hf download lance-format/lerobot-pusht-lance --repo-type dataset --local-dir ./lerobot-pusht

Then point Lance or LanceDB at ./lerobot-pusht/data.

Search

PushT does not ship a vector index out of the box — observation states are low-dimensional and most robotics workflows look up by index rather than by similarity. The bundled identifier columns (episode_index, task_index, frame_index) make exact lookups a single filtered scan. The example below pulls the first few frames of episode 0 from the frames table.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/lerobot-pusht-lance/data")
frames = db.open_table("frames")

slice_ = (
    frames.search()
    .where("episode_index = 0 AND frame_index < 10", prefilter=True)
    .select(["episode_index", "frame_index", "timestamp", "action", "observation_state"])
    .limit(10)
    .to_list()
)
for r in slice_:
    print(r["frame_index"], r["timestamp"], r["action"])

For similarity-style search across states or actions, attach an embedding column via Evolve and build an IVF_PQ index on it (see Evolve below). For visual similarity over rendered frames, the pre-extracted-frames pattern in Train below produces a table that can carry a learned image embedding alongside the pixels.

Curate

A typical curation pass for a robotics workflow starts with an episode-level filter — pick episodes with a particular task, length, or initial condition — and then drops down to the frames within those episodes. Stacking predicates inside a single filtered scan keeps the result small and explicit, and the bounded .limit(...) makes it cheap to inspect.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/lerobot-pusht-lance/data")
episodes = db.open_table("episodes")
frames   = db.open_table("frames")

# Pick a handful of episodes for the default task.
ep_rows = (
    episodes.search()
    .where("task_index = 0", prefilter=True)
    .select(["episode_index", "fps", "observation_images_image_from_timestamp",
             "observation_images_image_to_timestamp"])
    .limit(10)
    .with_row_id(True)
    .to_list()
)
ep_ids = [r["episode_index"] for r in ep_rows]

# Pull the frames belonging to those episodes for the next stage.
frame_rows = (
    frames.search()
    .where(f"episode_index IN ({', '.join(map(str, ep_ids))})", prefilter=True)
    .select(["episode_index", "frame_index", "timestamp", "action", "observation_state"])
    .limit(2000)
    .to_list()
)
print(f"{len(ep_rows)} episodes, {len(frame_rows)} frames selected")

Neither scan reads any video bytes. The MP4 segments live in the blob-encoded _video_blob columns and stay on disk until something explicitly asks for them.

Evolve

Lance stores each column independently, so a new column can be appended without rewriting the existing data. The lightest form is a SQL expression: derive the new column from columns that already exist, and Lance computes it once and persists it. The example below adds an action_magnitude and a large_action flag to the frames table, either of which can then be used directly in where clauses.

Note: Mutations require a local copy of the dataset, since the Hub mount is read-only. See the Materialize-a-subset section at the end of this card for a streaming pattern that downloads only the rows and columns you need.

import lancedb

db = lancedb.connect("./lerobot-pusht/data")  # local copy required for writes
frames = db.open_table("frames")

frames.add_columns({
    "action_magnitude": "SQRT(action[1] * action[1] + action[2] * action[2])",
    "large_action": "SQRT(action[1] * action[1] + action[2] * action[2]) > 5.0",
})

If the values you want to attach already live in another table (offline reward labels, classifier predictions, learned observation embeddings), merge them in by joining on the appropriate key — index for frames or episode_index for episodes:

import pyarrow as pa

rewards = pa.table({
    "index": pa.array([0, 1, 2]),
    "reward_to_go": pa.array([1.4, 1.3, 1.2]),
})
frames.merge(rewards, on="index")

The original columns and the inline video blobs are untouched, so existing code that does not reference the new columns continues to work unchanged. For column values that require a Python computation (e.g., running a visual encoder over the decoded video frames), Lance provides a batch-UDF API — see the Lance data evolution docs.

Train

A common pattern for vision-conditioned policy training is to pre-extract decoded frame pixels once into a derived LanceDB table — one row per frame, with the per-frame action and observation_state already joined in — and train against that table with the regular projection-based dataloader. take_blobs is the mechanism that makes the extraction step tractable: each episode's MP4 segment is randomly addressable in episodes.lance (the from_timestamp / to_timestamp columns give the segment bounds), so the pass can subset bytes on demand and write decoded frames into a fresh table without an external file store. Other workflows project the _video_blob columns from episodes.lance directly and decode at the batch boundary, or skip pixels entirely and train a state-only policy on frames.lance — the right shape is workload-specific. The actual training loop is the same Permutation.identity(tbl).select_columns(...) snippet in every case; only the source table and the column list change.

For a state-only policy, the frames table is already in the right shape — no pre-extraction needed:

import lancedb
from lancedb.permutation import Permutation
from torch.utils.data import DataLoader

db = lancedb.connect("hf://datasets/lance-format/lerobot-pusht-lance/data")
frames = db.open_table("frames")

train_ds = Permutation.identity(frames).select_columns(["observation_state", "action"])
loader = DataLoader(train_ds, batch_size=128, shuffle=True, num_workers=4)

For a vision-conditioned policy, train against a pre-extracted frames-with-pixels table that joins each frame's decoded image to its action and observation_state:

import lancedb
from lancedb.permutation import Permutation
from torch.utils.data import DataLoader

db = lancedb.connect("./lerobot-pusht-frames")   # local table produced by the one-time extraction
tbl = db.open_table("train")

train_ds = Permutation.identity(tbl).select_columns(["image", "observation_state", "action"])
loader = DataLoader(train_ds, batch_size=64, shuffle=True, num_workers=4)

The inline _video_blob storage and take_blobs still earn their place outside of the training loop — visualizing an episode in a notebook, sampling for human review, one-off evaluation against a held-out task, and the pre-extraction step itself — but they are not the dataloader.

Versioning

Every mutation to a Lance table, whether it adds a column, merges labels, or builds an index, commits a new version. Each of frames, episodes, and videos is versioned independently, so a column added to frames does not bump the version of episodes. You can list versions and inspect the history directly from the Hub copy; creating new tags requires a local copy since tags are writes.

import lancedb

db = lancedb.connect("hf://datasets/lance-format/lerobot-pusht-lance/data")
frames = db.open_table("frames")

print("frames version:", frames.version)
print("history:", frames.list_versions())
print("tags:", frames.tags.list())

Once you have a local copy, tag the table for reproducibility:

local_db = lancedb.connect("./lerobot-pusht/data")
local_frames = local_db.open_table("frames")
local_frames.tags.create("pusht-v1", local_frames.version)

Reopen by tag or by version number against either the Hub copy or a local one:

frames_v1 = db.open_table("frames", version="pusht-v1")
frames_v5 = db.open_table("frames", version=5)

Pinning supports two workflows. A policy locked to pusht-v1 keeps reproducing the same behavior while the dataset evolves in parallel. A training experiment pinned to the same tag can be rerun later against the exact same frames, so changes in metrics reflect model changes rather than data drift.

Materialize a subset

Reads from the Hub are lazy, so exploratory queries only transfer the columns and row groups they touch. Mutating operations (Evolve, tag creation, index builds) need a writable backing store, and a training pipeline benefits from a local copy with fast random access into the video blobs. Both can be served by a subset of the dataset rather than the full corpus. The pattern is to stream a filtered query through .to_batches() into a new local table; only the projected columns and matching row groups cross the wire, and the bytes never fully materialize in Python memory.

import lancedb

remote_db = lancedb.connect("hf://datasets/lance-format/lerobot-pusht-lance/data")
remote_frames = remote_db.open_table("frames")

batches = (
    remote_frames.search()
    .where("task_index = 0 AND episode_index < 50")
    .select(["episode_index", "frame_index", "index", "timestamp", "action", "observation_state"])
    .to_batches()
)

local_db = lancedb.connect("./pusht-task0-subset")
local_db.create_table("frames", batches)

The resulting ./pusht-task0-subset is a first-class LanceDB database. Every snippet in the Evolve, Train, and Versioning sections above works against it by swapping hf://datasets/lance-format/lerobot-pusht-lance/data for ./pusht-task0-subset. The same pattern applies to episodes and videos — narrow each table to the rows your workload needs, and the resulting database stays small enough to index and iterate cheaply.

Source & license

Converted from lerobot/pusht (LeRobot v3.0 dataset format). PushT is released under the Apache 2.0 license by the LeRobot project and the Diffusion Policy authors.

Citation

@misc{cadene2024lerobot,
  title={LeRobot: State-of-the-art Machine Learning for Real-World Robotics in PyTorch},
  author={R{\'e}mi Cadene and Simon Alibert and Alexander Soare and Quentin Gallou{\'e}dec and Adil Zouitine and Steven Palma and Pepijn Kooijmans and Michel Aractingi and Mustafa Shukor and Martino Russi and Francesco Capuano and Caroline Pascal and Jade Choghari and Jess Moss and Thomas Wolf},
  year={2024},
  url={https://github.com/huggingface/lerobot}
}

@inproceedings{chi2023diffusion,
  title={Diffusion Policy: Visuomotor Policy Learning via Action Diffusion},
  author={Chi, Cheng and Feng, Siyuan and Du, Yilun and Xu, Zhenjia and Cousineau, Eric and Burchfiel, Benjamin and Song, Shuran},
  booktitle={Robotics: Science and Systems},
  year={2023}
}