Voxel Car Demo

Random procedural generation plus closed-loop model evaluation. Closed-loop mode runs OccNet occupancy prediction with either standard A* planning or a pretrained PPO-RL controller.

Default model repo: mmkuznecov/SynthOccPredModels

Generate a procedural voxel road scene, preview camera frustums, render a side-by-side video, or save a dataset sample.

World

Seed

0 1000000

Grid size (voxels / side)

40 140

Max obstacle height (voxels)

5 22

Road half-width (voxels)

2 6

Noise scale

8 40

Shoulder extra radius (voxels)

0 8

Trajectory

Number of segments

2 12

Per-waypoint noise amplitude

0 1.5

(will populate after preview)

Drive / render

Number of frames

10 120

Camera image width (px)

96 320

Camera image height (px)

72 240

Output FPS

5 24

Cameras

Offsets are in metres from car centre. +right = passenger-side. +yaw = rotate camera toward car's right.

Frustum colour on BEV

Draw this camera's frustum / save its video

Forward offset (m)

-5 5

Right offset (m)

-3 3

Height (m)

0.3 6

Yaw vs car forward (deg)

-180 180

Horizontal FOV (deg)

30 140

Camera shown in composed video

Dataset save

Output root directory

Save status

Layout preview — BEV at start pose

Aligned BEV + selected camera view