README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/sensor_frames.parquet
license: other
tags:
  - robotics
  - manipulation
  - multimodal
  - tactile
  - lance

intelligencefactoryy/umi_table02

Multi-sensor manipulation data from the UMI multi-sensor rig (D405 RGBD, Logitech C922 context camera, BNO085 IMU, PAXINI tactile pads), stored as a Lance table. One row = one D405 depth frame (~8 Hz reference clock).

• table: sensor_frames — 42925 rows x 70 columns
• sessions: 1
• episodes: 210

This dataset has episode sessions: rows carry episode_dir (e.g. "episode_047") alongside session_id. The *_frame_idx columns are only meaningful within their own episode's video — always join on (session_id, episode_dir) to find the right file, never session_id alone. Legacy rows (whole-session recordings, no episode gating) have episode_dir = None and use the flat assets/videos/<session_id>/ path.

Layout

Lance_Table/                 # the Lance table (open with lancedb)
assets/videos/session_2026-07-06_16-48-46/episode_210/d405/rec_color.mkv
assets/videos/session_2026-07-06_16-48-46/episode_210/d405/rec_depth.mkv
assets/videos/session_2026-07-06_16-48-46/episode_210/logitech/rec.mkv

Camera frames are not stored in the table — only frame indices (d405_color_frame_idx, d405_depth_frame_idx, logitech_frame_idx). Seek into the matching MKV using the index, joined on (session_id, episode_dir). Mapping:

column	video
d405_color_frame_idx	assets/videos/<session_id>/<episode_dir>/d405/rec_color.mkv
d405_depth_frame_idx	assets/videos/<session_id>/<episode_dir>/d405/rec_depth.mkv
logitech_frame_idx	assets/videos/<session_id>/<episode_dir>/logitech/rec.mkv

(drop the <episode_dir> segment for legacy rows where it's None.)

Usage

from huggingface_hub import snapshot_download
import lancedb

local = snapshot_download("intelligencefactoryy/umi_table02", repo_type="dataset")
t = lancedb.connect(f"{local}/Lance_Table").open_table("sensor_frames")
df = t.to_pandas()

# decode the matching frame
import av
row = df.iloc[0]
episode_sub = f"/{row.episode_dir}" if row.get("episode_dir") else ""
path = f"{local}/assets/videos/{row.session_id}{episode_sub}/logitech/rec.mkv"
with av.open(path) as c:
    for i, frame in enumerate(c.decode(c.streams.video[0])):
        if i == int(row.logitech_frame_idx):
            img = frame.to_ndarray(format="rgb24")
            break