kdaterao/so101_locate_gripper

SO-101 gripper point labels for visual point grounding, with legacy bounding-box labels retained.

Labels	0
Point labels	1129
Images	1201

Layout

images/           # JPEG frames (epXXXXXX_fXXXXXX_{top,side}.jpg)
point_labels.jsonl # user-clicked gripper points (pixels + normalized coordinates)
labels.jsonl       # legacy click-drag boxes (xyxy pixels), when present

Point label row schema

{
  "image": "images/ep000500_f000120_top.jpg",
  "phrase": "SO-101 gripper",
  "point_xy": [x, y],
  "point_xy_norm": [x_normalized, y_normalized],
  "point_target": "center_of_gripper_jaws",
  "width": 640,
  "height": 480,
  "episode": 500,
  "frame": 120,
  "camera": "top"
}

Points are collected with a single click using src/robot101/data/collect_points.py. Point labels are intended for the custom MolmoPoint fine-tuning workflow.

python -m robot101.data.collect_points batch