so101_recovery_2task — SO-101 cup repositioning as two tasks, 749 episodes

Cup start positions for the 394 pick episodes (left) and start → placed transport for 350 of the 355 place episodes (right) — one color per operator, anonymized.

Front camera (fixed, elevated view), 8× speed — pick episodes followed by place episodes.

749 teleoperated demonstrations on the SO-101 arm, collected by 8 operators and recorded as two separately instructed tasks rather than one continuous motion:

"pick up the cup near the blue circle" — 394 episodes · 89,269 frames

"place the cup on the blue circle" — 355 episodes · 61,986 frames

This dataset is part of a larger coffee-making robot project. Splitting the cup-repositioning primitive into two instructions is what lets a policy be commanded mid-sequence — pick alone, place alone, or the two chained — instead of only replaying the whole motion.

This dataset was created using LeRobot.

Quick facts

Episodes / frames	749 / 151,255 @ 30 fps
— pick	394 / 89,269
— place	355 / 61,986
Operators	8, teleoperating with an SO-101 leader arm
Cameras	front (fixed, elevated front view of the workspace), wrist (gripper-mounted) — both 640×480
Robot	SO-101 follower, 6-DoF
Format	LeRobotDataset v3.0

Scene geometry

• Two fixed cameras — positions never changed during collection, so pixel coordinates are consistent across all episodes.
• The blue circle is fixed to the table. The cup is the only object that moves between episodes.
• Most episodes use a red-banded paper cup; 71 pick episodes (18 %) use a plain white one.
• Background coffee machines are the deployment context of the parent project — kept exactly as-is in every frame.

Quality

Every episode was reviewed frame by frame before release. Recordings with a fault — a hand entering the frame, the arm never contacting the cup, or zero motor motion — were removed rather than patched. Camera keys and task strings were verified against the video content, not trusted from the recording config.

Trained models

Four π0.5 fine-tunes trained on this dataset, in the two-task collection: expert-only vs full fine-tune, at 5K and 10K steps.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("aailabkaist/so101_recovery_2task")
print(ds.num_episodes, ds.num_frames)
print(ds.meta.tasks)

---

A single-task variant of this setup — one instruction covering the whole pick-and-place motion, with a plain white cup — lives in the SO-101 · cup → blue circle (single-task) collection.