DeepRoute.ai unveils Inference Engine for autonomous vehicles

Shenzhen-based self-driving technology startup DeepRoute AI showcased innovative technologies at the 2021 Consumer Electronics Show (CES), including an inference engine, DeepRoute-Engine, its second-generation all-in-one sensing solution, and human-like planning.

“DeepRoute has made exciting progress over the past year, and we look forward to presenting our latest solutions to a world-class audience at CES,” said Nianqiu Liu, Vice President of DeepRoute.ai. “We’ve perfected our level 4 self-driving technology both in software and hardware, and have strengthened our partnerships with OEMs and ride-hailing companies, to continually build the future of smart mobility.”

At CES, the company debuted DeepRoute-Engine, an inference engine that speeds up the neural network computation, allowing algorithms to run on an energy-efficient computing platform, the company said in a release. This results in a 6x faster inference process than that of widely used open source deep learning frameworks (e.g. TensorFlow, PyTorch, Caffe) and is compatible with GPU of various brands, including NVIDIA, Intel, and AMD, to greatly minimize the computation needed for level 4 self-driving technology.

Because of this powerful inference engine, DeepRoute’s level 4 self-driving technology can run on its 45-watt computing platform, DeepRoute-Tite, which is extremely power-efficient, consuming only 1/9 of the power, compared with traditional level 4 self-driving solutions.

Along with the debut of the inference engine, DeepRoute has also showcased its second-generation all-in-one sensing solution, DeepRoute-Sense II. It houses three LiDAR sensors, seven HDR cameras (DeepRoute-Vision), Global Navigation Satellite System (GNSS), inertial measurement unit (IMU), and a synchronization controller (DeepRoute-Syntric). This combination leads to an accurate object detection range of 140m (153 yards), the startup claims. Notably, the compact roof-box design minimizes the additional drag to only 7%, a huge decrease from the 20% on traditional sensing solutions, allowing for low energy consumption.

DeepRoute also highlighted its cutting-edge human-like planning at CES through a video of a fully autonomous road test in Shenzhen’s heavy traffic. “We call it ‘human-like’ because rather than pre-planned paths, our vehicle adopts a real-time planning method, just like a human driver who constantly thinks about how to drive safely and efficiently. This human-like planning enables a car to smoothly pass by an unplanned construction zone and choose the fastest lane through perceived traffic dynamics,” said Nianqiu Liu.