A LiDAR fires 120,000 measurements ten times a second. Keeping all of it is too slow; discarding it loses the kerb in front of you. We keep 5 cm detail where the vehicle will be in two seconds and coarsen it further out — 331× less memory, 100% of points still accounted for, 32 ms per scan on a laptop CPU.
Not a video. The Grid Lab recomputes 48,000 cells in your own browser as you move the sliders — the same algorithm, ported and tested against the Python line by line.
Surface depressions found per scan. Identical code, identical thresholds, five drives across three sensors. The only thing that changes is the ground.
| drive | sensor | surface | depressions / scan |
|---|---|---|---|
| Singapore, after rain | 32-beam | wet tarmac | 0 |
| Singapore, night | 32-beam | arterial road | 6 |
| Boston | 32-beam | city street | 36 |
| Karlsruhe (KITTI) | 64-beam | residential | 210 |
| Garching, off-road | 128-beam | gravel, ruts, mud | 1,122 |
We say depression, not pothole. We checked: only 1 of 44 sat on class road — the rest are kerbs and verges. The geometry is real and correctly located; calling them all potholes would be overclaiming.
Three sensors, three countries, paved and unpaved. The same code runs all five.
It reports its height, its size, how many points it holds and what they are.
In the Grid Lab the scan re-bins in your browser and the numbers change with it.
About 120,000 points per sweep from a spinning LiDAR.
SalsaNext labels every point: road, building, car, person… 20 classes.
20 classes collapse into three: drivable, static obstacle, dynamic object.
Cells grow with distance. Each keeps its height — that is the "2.5D".
Overhangs, surface defects, the reachable corridor, moving vs parked.
Slow down, steer around, or stop — the map becomes a decision.
| ring | radius | cell | why |
|---|---|---|---|
| 0 | 0 – 10 m | 5 cm | kerbs, potholes, feet |
| 1 | 10 – 25 m | 10 cm | vehicles, poles |
| 2 | 25 – 50 m | 25 cm | buildings, lanes |
| 3 | 50 – 100 m | 50 cm | layout only |
Every cell size is a whole multiple of 5 cm and every ring shares one origin, so the rings tile without seams. 70.7% of a scan's points fall inside 10 m; a uniform grid spends the same memory on the empty far field.
The recognition model was trained on 64-beam KITTI. On the 32-beam nuScenes drives its labels are noisier; the grid itself does not change — that is the point of running it.