MIT researchers create single deep neural network to support autonomous vehicles

Using a NVIDIA DRIVE AGX Pegasus, researchers at the Massachusetts Institute of Technology (MIT) have established a single deep neural network (DNN) for autonomous vehicles (AV).

Conventional end-to-end autonomous driving solutions typically rely on cameras, due to the prohibitively high memory and computation cost of processing 3D data.

Yet, researchers were able to make use of the NVIDIA hardware to rapidly process LiDAR, a remote sensing technology, data. To overcome the issues usually faced by processing the massive amounts of data generated from AV in real-time, MIT decided to approach the problem by using a DNN. With a number of improvements made to boost processing times and energy efficiency, the researchers say their system improves robustness and reduces the number of errors during events like sensor failures.

Unmapped locations provide an extremely difficult terrain for AVs to negation, as the machine backbone will need to perform many calculations to successfully navigate the surroundings. But the unique ability of the NVIDIA-supported system allows for data to be processed quickly even in situations such as this. The researchers published the results in a paper titled ‘Efficient and Robust LiDAR-Based End-to-End Navigation’ and fully laid out their methods to achieve this technological advance.

 “End-to-end learning has produced promising solutions for reactive or instantaneous control of autonomous vehicles directly from raw sensory data… Our framework has been evaluated on a full-scale autonomous vehicle and demonstrates lane-stable as well as navigation capabilities,” said the researchers in their paper.

With the size of the global autonomous car market forecast to reach more than $37 billion by 2023, according to data from Statista, there is clearly a demand for solutions that can reduce vehicle crashes and fatalities, at the same time as saving businesses money on personnel expenses.