Hugging Face's logo

Datasets:
Litian2002
/
tasl-fr3-10task-250ep-video

Tasks:
Robotics
Modalities:
Tabular
Time-series
Video
Formats:
parquet
Size:
10K - 100K
Tags:
LeRobot
VLA
franka
fr3
real-robot
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
tasl-fr3-10task-250ep-video / README.md
Litian2002's picture
Litian2002
README: merge task table into a task/episode index (eval id, episode ranges, frame counts, demo links)
4dd6d2d verified
preview code
|
Raw
History Blame Contribute Delete
6.57 kB
metadata 
license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - VLA
  - franka
  - fr3
  - real-robot
configs:
  - config_name: default
    data_files: data/*/*.parquet

TASL FR3 — 10-task real-robot manipulation dataset (video format)

Teleoperated (GELLO) demonstrations collected on a Franka FR3 + Robotiq gripper bench at TASL, in LeRobot v2.1 format. Used to LoRA fine-tune π0.5-DROID.

  • 250 episodes — exactly 25 per task, 10 tasks
  • 66,463 frames @ 15 fps (~74 minutes of teleoperation)
  • 2 camera views, 224×224 RGB, stored as h264 mp4 (500 videos)

This is the video-format copy, made so the LeRobot dataset visualizer can open it (that viewer rejects image-only datasets). The image-in-parquet original — which is what the policy was actually trained on — lives at Litian2002/tasl-fr3-10task-250ep. Frame data is identical; only the storage format differs.

Tasks & episode index

Episodes are grouped by task and numbered contiguously — task_index = k covers episode_{25k} … episode_{25k+24}, 25 demonstrations each.

eval id is the short code used by the evaluation harness and by the rollout datasets (e.g. T4-a-OOD1_r02_F). It is not stored in this dataset, but the two orderings line up one-to-one.

eval id task_index instruction episodes frames (median / total) watch first demo
T1-a 0 pick up the blue cup and place it into the red cup 0–24 264 / 6,788 episode_0
T1-b 1 stack the red block on top of the blue block 25–49 353 / 10,051 episode_25
T2-a 2 press the blue button 50–74 175 / 4,678 episode_50
T2-b 3 close the lid of the wooden shape sorter box 75–99 206 / 5,831 episode_75
T3-a 4 align the three colored blocks to the same orientation 100–124 193 / 5,099 episode_100
T3-b 5 rotate the red block so that it is perpendicular to the blue block 125–149 270 / 7,228 episode_125
T4-a 6 insert the orange block into the wooden shape sorter box 150–174 258 / 6,295 episode_150
T4-b 7 insert the book into the black book stand 175–199 259 / 7,602 episode_175
T5-a 8 pull the smaller book out of the black book stand 200–224 188 / 4,813 episode_200
T5-b 9 pull the small block out from under the large block 225–249 317 / 8,078 episode_225
all 10 tasks 250 66,463

File paths

Each episode is three files, indexed by the same zero-padded number:

videos/chunk-000/observation.images.exterior/episode_%06d.mp4   # table view (ZED 2i)
videos/chunk-000/observation.images.wrist/episode_%06d.mp4      # eye-in-hand (ZED Mini)
data/chunk-000/episode_%06d.parquet                             # state / actions / indices

So T4-a (insert the orange block) is episode_000150 … episode_000174.

Columns

column shape / dtype meaning
observation.images.exterior 224×224×3 video exterior camera (ZED 2i), looking at the table
observation.images.wrist 224×224×3 video wrist camera (ZED Mini, eye-in-hand)
state float32 [8] 7 joint angles + 1 gripper width
actions float32 [8] 7 joint velocities + 1 gripper command, normalized to [-1, 1]
task_index int64 index into meta/tasks.jsonl
episode_index, frame_index, index, timestamp standard LeRobot fields
done, is_success, intervene_flag bool written by the collection stack but constant — carries no information

Notes / caveats

  • No filtering was applied. Every frame recorded during teleoperation is present; no idle-frame removal, no success filtering. Idle frames are effectively absent anyway (only 0.4% of frames have max |joint velocity| < 0.05).
  • is_success and intervene_flag are True on every frame, so they cannot be used to select successful demonstrations.
  • Episode lengths: min 4, 1st pct 124, median 248, 99th pct ~700, max 971 frames. episode_index=172 is only 4 frames long (0.3 s) and should be dropped.
  • The two camera images are a centre-square crop of a 1280×720 ZED frame, resized to 224 — roughly 56% of the original horizontal field of view.
  • Collected 2026-08-16 and 2026-08-20.

Loading 

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("Litian2002/tasl-fr3-10task-250ep")

Fine-tuned checkpoints trained on this dataset: Litian2002/pi05-droid-franka-lora-10task