puheliang/lerobot_fmc3_grab_box_via_middle_v1

SO101 Tape Pick-and-Place via Middle Waypoint

Summary

This dataset contains real-world teleoperation demonstrations collected with a Seeed Studio SO101 setup using LeRobot. The task is:

1. Pick up a tape object.
2. Place it at an intermediate middle waypoint.
3. Move it from the middle waypoint into a box.

The dataset is intended for behavior cloning / fine-tuning of vision-language-action or visuomotor policies on a longer-horizon version of the original tape-to-box task.

Collection Setup

• Robot: so101_follower
• Teleoperation device: so101_leader
• Control modality: human teleoperation
• Cameras:
  • top: RGB, 640x480, 30 FPS
  • wrist: RGB, 640x480, 30 FPS
• Observation state dimension: 6 joint positions
• Action dimension: 6 joint targets
• Frame rate: 30 FPS

Current Local Dataset Status

• Local repo id: puheliang/lerobot_fmc3_grab_box_via_middle_v1
• Recorded episodes currently present in local metadata: 31
• Current episode index range: 0 to 30

Task Description

Compared with the earlier direct pick-and-place dataset, this version adds a mandatory intermediate waypoint before the final placement into the box. This makes the task longer-horizon and is useful for fine-tuning an existing grasp policy into a more structured multi-stage policy.

Intended Use

• Fine-tuning an existing pi0 or other LeRobot-compatible policy.
• Comparing direct placement vs. waypoint-conditioned placement behavior.
• Evaluating whether adding a middle waypoint improves motion structure and robustness.

Notes

• Demonstrations were recorded locally and may continue to grow over time.
• Before training, it is recommended to review episodes for human hand/body occlusions or other visual distractions in the camera views.
• The dataset includes synchronized robot state, action, and dual-camera video streams.