cubestack_64_flora

NVIDIA Flora annotations for 64 episodes of the ultra cube-stack dataset (two-arm robot, LeRobot v3.0, 50 fps): 32 stack cubes + 32 unstack cubes.

Episode selection

exit_type == success, not in meta/excluded_episodes.json, ≤ 119.5 s (Flora's validated limit), then 32 per task sampled at random (seed 0). Eligible pool: 447 stack, 408 unstack. Total 12.4 min of video. Episode ids (ep%06d) are the source dataset's episode_index; jobs_valid.jsonl also lists each episode's source_session.

Pipeline

Each episode's three cameras were cut out of the packed AV1 files and re-encoded to H.264 (x264 crf 16, native 50 fps, frame counts checked against the episode length). Cameras: main = observation.images.head, left = left_wrist, right = right_wrist. Robot profile yam (Flora's generic robot-arm profile).

Stage	Model	Output
TempSeg	nvidia/Flora-TempSeg @ 114a1076 (Qwen3.5-9B base @ c2022362)	segments/<id>.json
Captioner	nvidia/Flora-Captioner @ fef0bbfe	annotations/<id>.json (subtask segments + captions)
Rewarder	nvidia/Flora-Rewarder (vLLM, greedy)	rewards/dense.jsonl, rewards/terminal.jsonl

• Dense: 1–3 per 0.5 s (failure / suboptimal / normal), 16 s clips at 4 fps every 10 s. Collage order left, head, right.
• Terminal: 1–5 task-success score over up to 64 frames, with token probabilities and expected_score.
• No episode instruction was passed to the Captioner; the Rewarder's task text is the dataset's task string.

Caveat: rewards

All 64 episodes are recorded successes, yet every terminal score is 1 (mean expected score 1.23 stack, 1.13 unstack), and dense means are 2.84 (stack) / 1.49 (unstack). The Rewarder does not transfer as-is to this robot/camera setup and the two-word task strings; treat these reward files as a baseline, not ground truth.