license: apache-2.0
nyu center for robotics & embodied intelligence
lab for autonomous and intelligent robotics
project: embodied at sea (behavioral cloning for marine archaeology & conservation)
robot: blue robotics's "bluerov2"
contributors: Dr. Chris Clark (PI), Greta Perez-Haiek (Lead), Luca Macesanu, Kevin Angers, Timothy Gambin (Advisor)
environment: rectangular swimming pool at NYU CREO, filled with chlorinated water at about a little over a foot's depth of water.
factors: neutral background (shades of blue hexagons with white rimming), neutral lighting (well-lit and constant), disturbances & distractors (variable)
Snapshot date: August 18, 2026
This folder contains 8 LeRobot-format (codebase_version v3.0) datasets recorded
on a BlueROV2 performing rod-grabbing tasks in a pool, with and without
simulated currents. Each dataset is packaged as a .zip containing the
standard LeRobot layout (data/, meta/, videos/).
Counts below are taken from each dataset's meta/info.json.
- Episodes =
total_episodes(number of recorded episode/trajectory runs) - Frames =
total_frames(total timesteps across all episodes, fps = 50.0) - Trajectories =
total_successes(episodes that ended in a successful grab)
| Dataset | Episodes | Frames | Trajectories (successes) |
|---|---|---|---|
| grab green rod | 101 | 26,774 | 101 |
| grab green rod under currents | 110 | 37,982 | 101 |
| grab red rod | 101 | 23,550 | 101 |
| grab red rod under currents | 102 | 30,171 | 101 |
| grab white rod | 102 | 26,853 | 100 |
| grab white rod under currents | 101 | 27,934 | 101 |
| grab yellow rod | 101 | 22,628 | 101 |
| grab yellow rod under currents | 102 | 28,220 | 101 |
Notes
- "Episodes" counts every recorded run, including any that did not end in a
successful grasp; "Trajectories" (
total_successes) counts only the successful ones. In most of these datasets nearly all episodes succeeded (only "grab white rod" has 2 non-successful episodes: 102 episodes vs 100 successes). - Frame count varies with episode length/duration, not just episode count — e.g. "grab green rod under currents" has the most episodes (110) and by far the most frames (37,982), consistent with currents making the task take longer per episode.
robot_type: bluerov2, single task per dataset (total_tasks= 1).