Deep learning-powered autonomous truck perceives objects at 1 km
Autonomous truck company TuSimple uses Amazon Web Services (AWS) technology to train their Level 4 self-driving fleet
TuSimple, a global autonomous truck technology leader, publicly demonstrated its Level 4 fully autonomous driving solution that it claims can see and interpret objects up to 1,000 metres using a camera-based solution. This is of note as autonomous vehicles have recently been found to struggle in imperfect conditions and to operate with complete safety on roads with human drivers, meaning greater detection range is highly desirable, especially for large commercial vehicles. The solution is also of note for its solution. Most autonomous vehicles so far have been highly reliant on radar and LIDAR, the latter of which uses lasers, and is relatively expensive.
Using this system, TuSimple's self-driving Peterbilt semi-trucks are currently hauling commercial cargo for revenue on a daily basis along the I-10 corridor.
These drives are accomplished using TuSimple's proprietary deep learning-based autonomous driving system. The system uses cameras as primary sensors, and sensor fusion technology to achieve a pixel-level interpretation of the visible environment 1,000 meters away with the ability to locate itself within four inches on the road at all times. This comprehensive view of the truck's environment as it drives supports the company's dock-to-dock business model where its trucks will navigate safely between highways and distribution centres.
To achieve the 1,000 meter perception range, TuSimple spent two years developing the deep learning algorithms that are the instructions used to help the perception system understand terabytes (TB) of data created per vehicle every day. TuSimple turned to Amazon Web Services (AWS) to support the company's development process, relying on AWS's machine learning expertise and nearly unlimited compute and storage capabilities. The data collected from the company's AV test vehicles is stored using AWS Snowball Edge devices, each capable of storing 100 TB of data, with onboard computational capacity that allows for local data analysis and data compression. The device is then shipped directly to an AWS data center for secure ingestion into the AWS Cloud. Once the data is uploaded to the AWS Cloud, TuSimple's development engineers use Machine Learning on AWS and Amazon Elastic Compute Cloud (Amazon EC2) P3 high performance compute instances which are ideal for the computationally-advanced machine learning workloads like autonomous driving.
"Intensive training of autonomous vehicle software is critical to the success of all autonomous vehicles but even more so for the Class 8 tractor-trailers that haul freight 24/7 on our highways. It can take a football field, or 100 meters, to stop a fully loaded truck travelling at highway speeds of 65 miles per hour which means our software must look into the future, recognize and makes informed decisions about any situation it encounters," stated Dr. Xiaodi Hou, co-founder and CTO, TuSimple. "AWS and its data transfer service is the perfect collaborator for us with near infinite compute and storage capabilities allowing us to test the widest range of simulated driving experiences. In our pursuit of autonomous driving perfection, we generated 20 million miles of testing and terabytes a day testing our software on AWS's platform to ensure our system makes the most informed and safest decisions possible while driving."
AWS also provides a robust simulation platform for TuSimple to run millions of simulated miles on each of its algorithms by which the truck is guided. Then, TuSimple tests the software on the vehicle on a closed track, before driving it on the public highway. TuSimple trains professional CDL-holding truck drivers to monitor the vehicle and has a test engineer monitoring the system and the driver on all test runs. In 2019, TuSimple will deploy 200 trucks in the United States and it claims that it will be the world's largest AV truck fleet in commercial transportation service by that time.
"Camera-based autonomous systems rely heavily on deep learning models to make self-driving possible. TuSimple leverages the machine learning capabilities on AWS to create and train their models which reduces training time from days to hours," said Ryan Gavin, Head of Artificial Intelligence and Machine Learning Marketing for AWS. "This process is faster and more efficient than before, and just one example of how AWS is helping to improve the efficiency, cost and speed of development as TuSimple works to make safe and reliable autonomous vehicles a reality."