license: cc-by-4.0
pretty_name: DrivingBench Simulator
tags:
- gaussian-splatting
- autonomous-driving
- simulation
DrivingBench Simulator
The Gaussian splats behind the DrivingBench simulator, a drivable twin of
the v1.1 cone course. They were trained on the Tesla Model 3's own road camera on 27 October 2026,
from frames taken between 15:30 and 17:00. The simulator itself (the course, its cones, the
drivable region and the car fitted to the Tesla's logs) is in the
DrivingBench harness, under sim/.
| File | Gaussians | Size | Used by | SHA-256 |
|---|---|---|---|---|
oct27-nocones-full.splat |
2,924,873 | 93.6 MB | the simulated road camera | c306e7ae33aaa002c138b070ab31710a68827e8a20f0db1610e933a352a17c9c |
oct27-nocones.splat |
1,500,000, the most opaque | 48 MB | the lot explorer in a browser | 672709eddd03b05dc93e1f127b21887e31fdd6c5aa2b931ddc16a4ac34d6fa17 |
Both are .splat files (32 bytes per Gaussian: position, scale, colour and opacity, rotation), in
the course frame in metres. The real cones and the haze along the car's path are removed; the
simulator places the course's cones back as objects.
Use
From a checkout of the harness at v1.1:
uvx --from huggingface_hub hf download drivingbench/simulator --repo-type dataset \
--include "*.splat" --local-dir runs/sim/explorer
uv run python -m drivingbench.sim.runtime # the simulator, on 127.0.0.1:8877
uv run python -m drivingbench.sim.explorer.serve # the lot explorer, on 127.0.0.1:8890
The harness's docs/sim.md describes how the splats were built and how faithful the simulator is.
License
CC BY 4.0. If you use the splats, please link drivingbench.com.