chans/quantgate-so101-coaster-positions-v2

QuantGate SO-101 demonstrations: yellow can and far start positions

These are 50 teleoperated demonstrations of the same task and setup as chans/quantgate-so101-coaster. They were recorded for QuantGate's pre-registered fine-tune v2 (amendment 9), which targets the two conditions where the first position fine-tune showed no gain: a smaller yellow can, and a cola can started 6 cm off the mark.

• Task: pick up the can and place it upright on the coaster.
• Conditions (seeded order; start_conditions.csv maps episode_index to object and start_offset_cm):
  • yellow can (190 g, 50 mm): 5 episodes at each of −4, −2, 0, +2 and +4 cm;
  • cola can (268 g): 13 at +5 cm and 12 at +7 cm.
  • Held out on purpose: the yellow can at +3 cm and the cola can at +6 cm are the evaluation conditions and do not appear here.
• Episodes: 50. Frames: 37,227 at 30 fps (743–746 frames each, 25 s).
• Robot and cameras: as chans/quantgate-so101-coaster (SO-101 follower and leader, overhead C922 640×480, OpenMV wrist camera 320×200, AV1 video). The gripper has rubber bands on the fingertips; the wrist-camera mount was re-aimed and fixed before recording.
• All episodes are successful demonstrations. checksums.sha256 covers every file.

Use

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("chans/quantgate-so101-coaster-positions-v2")

People in the data

The operator's hand and arm can appear in the overhead view while they reset the scene. No faces are recorded, and there is no audio.

Related

• chans/quantgate-so101-coaster (120 demos) and chans/quantgate-so101-coaster-positions (50 demos, −4…+4 cm).
• Code, pre-registrations and results: the QuantGate repository (public from 2026-10-26).

License: CC BY 4.0.