DeepRoute.ai has announced plans to develop end-to-end neural network software for self-driving, writes Nick Flaherty.

The software will run on the latest DRIVE Thor platform from Nvidia, which is based on the new Blackwell GPU architecture, designed for transformer and generative AI workloads. Scheduled for launch next year, it will be adopted by passenger vehicles first, integrated with DeepRoute.ai’s smart driving software. 

Unlike traditional autonomous driving solutions that separate into perception, prediction, planning and control modules, the end-to-end model integrates all four into one neural network, so no information is lost.

The model learns from a repository of video clips that provide insight into more nuanced driving behaviour. One example of this sophistication is the model’s capacity to understand courtesy and consideration towards pedestrians during heavy rain, adjusting its actions to take cues from their behaviour. 

“The end-to-end model is revolutionary for autonomous driving, signalling the end of rule-based AI and the dawn of a new era that is learning-based,” said Maxwell Zhou, CEO of DeepRoute.ai.

The company began conducting road tests of the end-to-end architecture from training to deployment on the current processor chip, called Orin, in August 2023.

Road tests demonstrated proficiency in navigating complex scenarios with oncoming vehicles and pedestrians in a natural, polite manner while ensuring safe and efficient driving. Another example showed the vehicle carefully changing lanes to pass when recognising a taxi dropping off passengers.

“DeepRoute has demonstrated remarkable technology innovation prowess in recent years, evolving into a staple within the transportation industry for AI driving systems,” said Rishi Dhall, vice-president of automotive at Nvidia.

Deeproute.ai continues to add data for the next iteration of the model.