description: '5,172 RGB frames (1 fps, 640 px) decoded from the camera streams of
  75 DROID robot-arm episodes, for the Nebius x Voxel51 Physical AI workshop. Every
  frame keeps its episode_id, camera and timestamp plus the robot state at that instant
  (gripper_open, joint_speed_norm, ee_linear_speed), and carries: CLIP embeddings
  with similarity (frames_sim) and UMAP (frames_viz) brain runs, uniqueness, near-duplicate
  tags, an episode-level train/val split, Grounding DINO open-vocabulary auto-labels
  (auto_labels), YOLO11n predictions from a Nebius Serverless AI Jobs fine-tune (yolo11n_preds)
  and the evaluation run eval_yolo. Derived from DROID (https://droid-dataset.github.io),
  CC-BY-4.0.'
fiftyone:
  version: '>=1.21.0'
format: FiftyOneDataset
license: cc-by-4.0
name: droid-frames
tags:
- droid
- fiftyone
- image
- manipulation
- object-detection
- object-detection
- robotics
- workshop