Serving the World's Largest Emerging Automobile Market
Home > Feature > China leading the world in vision-based autonomous driving, here’s how AutoX is doing it

China leading the world in vision-based autonomous driving, here’s how AutoX is doing it

[ UPDATE ] AutoX just gave the first Robotaxi (xTaxi) ride in California this week, transporting the first Robotaxi passenger in the State,  Angie, to Safeway for her grocery shopping, the company said in a Medium post. It recently launched the first ever Robotaxi pilot service in California, the xTaxi, and began accepting early rider applications after becoming the second company in the world to be permitted by the California Public Utilities Commission to operate a Robotaxi pilot program.

Two months ago Tesla sponsored its first Autonomy Day where Elon Musk expounded at length about how LiDAR is not needed for autonomous driving. Most Chinese companies working on autonomous driving, including Baidu’s Apollo, Pony.AI and, believe LiDAR is essential part of the sensor suite. The media immediately pounced on Tesla, incorrectly stating that Tesla is at odds with “the approach of every major autonomous vehicle program.”

Tesla’s main argument is that it is essential to solve computer vision in order to achieve true Level 5 autonomous driving. Once vision is solved Musk believes LiDAR is no longer necessary, or even desirable. There are five major vision-based AV developers that are not dependent on LiDAR for their autonomous driving solutions:

Tesla (USA), (USA), MobilEye (Israel, USA), AutoX (China, USA), AIMotive (Hungary, USA, Japan) and Nullmax (China, USA).

AutoX Technologies, Inc. was founded in 2016 by Dr. Xiao Jianxiong, a former assistant professor of Princeton University commonly known as “Professor X” to his American students. Professor Xiao set up AutoX in San Jose, California with $9 million in seed funding. In late 2018 Dr. Xiao moved the company’s headquarters to Hong Kong where he could take advantage of the resources of his alma mater, the Hong Kong University of Science and Technology. Months later he set up a subsidiary across the border in Shenzhen and expanded into mainland China.

AutoX has been offering grocery delivery services in San Jose since August 2018. After downloading an app, users can browse and order vegetables and fruits which are delivered using an AutoX autonomous vehicle. Today AutoX delivery cars include safety drivers, but AutoX has developed a monitoring system similar to the wireless driving system offered by Phantom Auto to allow their vehicles to travel without safety drivers.

AutoX is operating the grocery delivery service in an attempt to gather driving data while making a profit on deliveries. Aside from its vehicles in San Jose, AutoX is also doing vehicle testing in Shenzhen and in Changzhou, Jiangsu Province. Driving data is used to train the neural networks which control AutoX artificial intelligence.

AutoX has achieved impressive success with automotive OEMs, building prototype and development vehicles for Ford, SAIC, BYD and Dongfeng. Earlier this year AutoX was able to parley its success with Dongfeng into an undisclosed investment above $10 million. Other AutoX investors include SAIC Capital and silicon chip giant MediaTek.

Many in China’s AV industry say tech giant Baidu’s Apollo autonomous driving platform is clearly in the lead. Baidu has alliances with more than 135 OEMs and over 90 percent of all autonomous vehicles being tested on Chinese roads include the Apollo autonomous software stack.

AutoX received an Autonomous Vehicle Pilot Permit from the California Public Utilities Commission (CPUC) on June 18, 2019, becoming the second of three companies (the other two being Zoox and Pony.AI) to have received such permits to offer robo-taxi services to the general public (with safety drivers) in California. AutoX also became one of six companies to have received permits in Guangzhou on June 20 for public road testing of intelligent connected vehicles (ICV) in the city.

From a technology standpoint AutoX may have quietly left Baidu and the other Chinese AV competitors behind, which accounts for China’s number one BYD and number two SAIC both signing development agreements with AutoX.

The AutoX autonomous driving software stack has three fundamental pillars: sensation, localization and planning.

Sensation builds an environmental model using cameras, radar, LiDAR, IMU, GPS, ultrasonic sensors and other devices that together determine the vehicle’s location, direction and speed, the surrounding environment, the players in the surrounding environment and what the players are doing.

Like Tesla, AutoX uses deep supervised learning through neural networks to turn 2D camera images into 3D models. AutoX first demonstrated a vision-based autonomous vehicle way back in 2016 and is one of only 5 autonomous driving software stack developers focusing on this strategy. LiDAR was added to their sensor suite later to meet OEM customer demands for improved redundancy. AutoX could get rid of LiDAR tomorrow and their autonomous driving solution would still function.

Baidu’s Apollo is fundamentally different: Apollo requires LiDAR data to create a 3D model on 2D camera images. Camera images and LiDAR results are aligned together so that LiDAR data can add 3D characteristics to the 2D camera images.

The problem is LiDAR leads to a local maximum where its capabilities are fundamentally unable to provide the level of visual perception demanded by autonomous driving. Autonomous driving system developers depending upon LiDAR are able to get their vehicles on the road very quickly but will eventually find themselves stuck at a point where their systems are very good, but unable to exceed the safety threshold where they have plateaued. LiDAR’s safety threshold is severely limited by its poor vertical resolution: the average LiDAR has between 16 and 128 vertical beams. Once an autonomous driving system reaches the point where vertical resolution becomes a limiting factor the vision problem must be solved to improve functionality further. AutoX technology is far superior to the Apollo camera image + LiDAR 3D texture kluge.

The second pillar of autonomous driving, localization, also known as path mapping, detects and comprehends the free space where the automobile can move, where all the drivable paths are, along with the lane markings, boundaries and the types of boundaries, lane boundaries even when there are no lane markings, road markings, traffic signs, lane splits, lane merges and destinations associated with each drivable path.

Unlike Tesla, AutoX uses high-resolution maps in a similar fashion as Baidu, Waymo and Cruise. AutoX has the ability to drive without maps and will do so when the system detects road construction or other circumstances where the high-definition maps are no longer reliable. In an area where road construction is underway AutoX has the capability to localize its surroundings on the fly, similar to Tesla’s localization functionality in their autonomous driving software stack. By using HD maps where the road conditions are already known, AutoX adds a layer of redundancy that Tesla lacks. Indeed, had Tesla included high-definition maps in its autonomous driving software stack, their system should have been able to avoid the accident that occurred on the morning of March 23, 2018, when a Tesla driving on AutoPilot steered directly into a concrete median, killing the driver in the resulting crash. The vehicle would have known the location of the concrete median and avoided driving straight into it. In this respect AutoX is better than Tesla.

The third pillar of autonomous driving, planning, sets the driving policy for the AV considering the vehicle, its surroundings and its interaction with other vehicles, pedestrians and animals in its environment. Driving policy plans for the future in a multi-agent scenario with multiple different behaviors exhibited by other agents. Planning uses reinforcement learning to observe states, take action and get rewards using recurring neural network learning.

Like Mobileye, AutoX uses recurring neural networks to determine driving policy during the planning stage of autonomous driving. According to Pan Jianwei, vice president of AutoX, the planning portion of the AutoX autonomous driving software stack is already feature complete, a stage Tesla does not plan to reach until the end of 2019. AutoX is presently in the application process to allow their vehicles to drive on California roads without safety drivers.

There are other players developing vision-based AV systems. There is only one company working with BYD, the world’s leading plug-in vehicle manufacturer, and SAIC, China’s second largest EV manufacturer. It will be interesting to watch which companies come out ahead in the race to autonomy.

AutoX is unquestionably a leader in the pack.

About the author:

Charlie Paglee is the CEO of Brannan Auto, an American automobile engineering and manufacturing company with multiple factories in China. Paglee is also the director of the China Autonomous Vehicle Experts group. Paglee has three decades of experience doing business in China and speaks fluent Chinese Mandarin. He can be reached via email at

*Note: The views expressed are those of the author and do not necessarily represent those of China Automotive Review. D

Leave a Reply