chans/quantgate-so101-coaster-positions

QuantGate SO-101 demonstrations with varied start positions

These are 50 teleoperated demonstrations of the same task and setup as chans/quantgate-so101-coaster. Here the can starts shifted along one axis. They were recorded for QuantGate's pre-registered position fine-tune.

• Task: "Pick up the cola can and place it upright on the coaster".
• Start offsets: −4, −2, 0, +2 and +4 cm from the original mark, 10 episodes each, in a seeded order. start_offsets.csv maps episode_index to start_offset_cm.
  • +3 cm is deliberately absent: it is the held-out evaluation position.
• Episodes: 50. Frames: 37,329 at 30 fps. Each episode runs the full 25 s window, 745–748 frames.
• Robot and cameras: the same as the companion dataset. SO-101 (Feetech STS3215), six joint positions, overhead 640×480 (C922) and wrist 320×200 (OpenMV AE3), AV1 video.
• Recording note: one empty episode (recording started by mistake) was removed with lerobot-edit-dataset delete_episodes. That position was then re-recorded in the same slot, so the order table still holds.
• checksums.sha256 covers every file.

How it was used

The fine-tune started from the G2 ACT checkpoint and ran 20k steps on these 50 episodes mixed with the 80 train episodes of the companion dataset (seed 1001). It was evaluated blinded on the real arm at +3 cm and at the original mark.

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.

License: CC BY 4.0.