1,469 drives through Bangalore, seven cameras each, 3.2 TB. We fine-tuned an end-to-end driving policy twice, once on the corpus as released and once on an edited version with 213 recordings removed, and ran both through the same ten closed-loop scenarios, 100 runs each. The edited version won.
A same-size random cut scored 58.9%, so the gain is from which recordings were removed, not from removing recordings. U-turn is the one scenario that slipped, and it is the one where standing-still footage was doing some work.
Edited minus full, with its 95% range. The pooled lift clears zero. No single scenario does at 100 runs, and the chart says so rather than rounding it away.
Measured from the release manifest. Every recording carries a timestamp, so the clock is the one label the corpus has. 1,431 of 1,469 drives were recorded between 11:00 and 17:00. Thirty-eight are after 18:00, none are at night, none are before 08:00. A policy trained on this corpus has never seen a headlight. 356 recordings are also missing the front-tilted camera, which the edit leaves alone since the other six are intact.
| Release | |
|---|---|
| recordings | 1,469 |
| cameras per recording | 7 |
| video | 1280×720 HEVC, 10 fps |
| GNSS with RTK | 60 Hz |
| recorded | May to Nov 2025 |
| median recording | 33 s |
| longest recording | 3 h 25 min |
| front-tilted camera missing | 356 |
| payload | 3.20 TB |
Every recording was scored on whether its cameras predict its own GNSS motion better than a shuffled pairing does. A drive where the car is not moving has nothing to predict, and a drive whose position fix has drifted has the wrong answer. 213 recordings scored at or near zero. 128 of them are a stationary car, at a signal, at a kerb, or behind something. That is not a bad recording, it is just not driving, and a policy trained on it learns to sit.
| Why it was removed | recordings | hours |
|---|---|---|
| stationary for more than 70% of the recording | 128 | 6.4 |
| RTK fix lost, position drifting | 41 | 2.9 |
| duplicate of another recording, same route, same minute | 22 | 1.1 |
| never leaves the depot | 17 | 1.3 |
| camera gap longer than one second | 5 | 0.2 |
| Removed | 213 | 11.9 |
| Kept | 1,256 | 77.9 |
Front camera, twelve seconds each. Six that the cut removed, two that it kept, for contrast.