DiffusionDrive
Truncated diffusion model for real-time end-to-end autonomous driving
0
- -Fast real-time decision-making: The model reduces diffusion denoising steps by 10 times, achieving faster decision-making speed.
- -High accuracy: In the NAVSIM benchmark test, the PDMS of Diffusion Drive is 3.5 times higher than the original diffusion strategy.
- -Diversity: The model has higher pattern diversity, scoring 64% higher than the original diffusion strategy.
- -Direct learning: The model can learn directly from human demonstrations without the need for additional training data.
- -High flexibility: Diffusion Drive can be easily integrated into in car sensor data and existing perception modules.
- -Easy to deploy: As a model, DiffuseDrive can be deployed on various autonomous driving platforms with good compatibility.
- -Open source: The code and model of the project will be open source for further research and development by the community.
Product Details
Diffusion Drive is a truncated diffusion model used for real-time end-to-end autonomous driving, which accelerates computation speed by reducing diffusion denoising steps while maintaining high accuracy and diversity. This model learns directly from human demonstrations and can achieve real-time autonomous driving decisions without complex preprocessing or post-processing steps. Diffusion Drive achieved a breakthrough of 88.1 PDMS in the NAVSIM benchmark test and was able to run at a speed of 45 FPS.
