Skip to content

DiffusionDrive

A truncated diffusion model for driving

Visit official site

What DiffusionDrive is

The model is described as a truncated diffusion policy that speeds up real time end to end driving decisions while keeping accuracy and diversity.

What you can do with it

  • Cut the number of denoising steps
  • Learn driving from human demonstrations
  • Fit into onboard sensor data

Application scenarios

  • City traffic is handled in real time
  • Highway following and overtaking
  • Turns and signals at intersections

Pros & cons

✓ What we like

  • Learns from human demonstrations directly
  • Runs at 45 FPS with 88.1 PDMS on NAVSIM

! What to watch out for

  • The model targets driving decisions
  • Integration with perception modules is assumed

FAQ

What benchmark figures are stated?

The material states 88.1 PDMS on NAVSIM and operation at 45 FPS.

What diversity figure is stated?

The feature list states a mode diversity score 64 percent above the original diffusion policy.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review