Pixel-J: Intelligent level to exceed Tesla generation
What can be the product of the combination of Pixel-J - one of the most car-making companies in China and the strongest AI company in China?
On June 8, Pixel-J, a joint venture between Baidu and Geely Automobile, held its first brand launch ROBODAY since its establishment a year and a half ago, officially releasing its first car robot concept ROBO-01. Based on this concept, Pixel-J's first production model will be officially unveiled this fall and officially launched in 2023.
The core technologies demonstrated in the ROBO-01, such as the automotive robot neural network (PIXEL-J Evolving Technology), true redundancy high-level autonomous driving solution, SOA-based cabin-driver fusion, and millisecond offline intelligent voice will also be implemented in the production car.
"We hope to build a smart car that is affordable to many people in terms of price, completely reaches the top level of current electric cars in terms of car performance and quality, but is far ahead of the generation in terms of intelligence." This is a big goal for me and the whole company," said Yiping Xia, CEO of Pixel-J, to CHNECAR. "
Specifically for the production version of the ROBO-01, Xia Yiping said, "To come up with a product in the SUV market that is fully ahead of and beyond the Tesla Model Y generation in terms of intelligence level."
How do achieve it?
01
The first mass production, armed to the teeth
The three mandatory courses in car robotics are defined by the set degree
"We (want) to make a truly intelligent benchmark car with high-level capabilities." In Xia Yiping's view, the excellent intelligent voice interaction, intelligent driving ability, and learning iteration ability of the Pixel-J car robot will be its core product competitiveness.
For this reason, ROBO-01's intelligent cockpit and intelligent driving system are armed to the teeth.
Pixel-J will debut Qualcomm's fourth-generation Snapdragon automotive digital cockpit platform-8295 chip in China. Compared with the Qualcomm 8155 chip, which has just been on board for a short time, the 8295 adopts a more advanced 5nm process (the 8155 adopts 7nm), and the computing power has been increased from 4TOPS to 30TOPS in one go.
The large computing power cockpit chip will support the intelligent cockpit and intelligent driving redundancy of the set degree.
ROBO-01's large integrated screen
The intelligent cockpit will achieve "millisecond" intelligent voice response and "full offline" intelligent voice function, which will get rid of the influence of network signals in tunnels, under bridges, parking lots, and other areas. In addition, its visual perception, voice recognition, lip capture, and other multimodal fusion of "human-like" interaction capabilities, human-machine interaction "natural and smooth".
At the same time, Qualcomm 8295 will also work with an exterior camera to form hardware redundancy for intelligent/autonomous driving functions. Xia Yiping called it the only model in the industry to achieve hardware redundancy.
It's worth noting that XPENG previously used a front camera and Infineon's computing platform as redundancy on the P7, with algorithms provided by Bosch. But by the P5, with the LIDAR on board, XPENG removed this redundancy.
Hardware solution for the first model of Pixel-J's autonomous driving system
The first model of Pixel-J will also be equipped with LIDAR, and with a concealable type.
The Pixel-J autonomous driving system is equipped with 2 NVIDIA Orin X chips, and the whole vehicle is equipped with 31 off-board sensors. This includes 2 lidars, 5-millimeter wave radars, 12 ultrasonic radars, and 12 cameras.
The ROBO-01 concept car's "jump light" LIDAR is located at the front hood, and the two LIDARs have a FOV of 180 degrees, which can better identify obstacles or pedestrians crossing left and right in driving scenarios such as "ghost probes" and left and right obstructions. When the intelligent driving mode is turned on, the two sets of "vision + dual LIDAR" autonomous driving systems not only keep independent operation but also can backup and cooperate with each other.
According to the official, this set of the automatic driving system has realized the "three domains of highway, city and parking integration", and has the point-to-point advanced automatic driving capability. The system has already passed the functions of an unprotected left turn, traffic light recognition, obstacle avoidance, and free on/off-ramp. When the product is delivered to the market, users can directly obtain the higher-order autonomous driving functions in multiple scenarios.
Xia Yiping told CHNECAR that compared to other brands whose smart/autonomous driving capabilities are limited by high precision map coverage and need to be landed city by city, the cities and scenarios that can be adapted when the production models of Pixel-J are landed are much broader. "Because of the greater reliance on machine learning (rather than heavy reliance on high precision maps), we have a very high level of generalization for city coverage. The quest is to (directly) give users a car-to-car smart driving experience."
"We want to come up with a product in the SUV market that is fully ahead and beyond the Tesla Model Y generation in terms of intelligence level." Xia Yiping stressed.
02
Super Model Y with Baidu
Established less than two years ago, Pixel-J 's first product is going to challenge Tesla in terms of intelligence. I'm afraid this is not a "delusion" to attract attention. Don't forget that one of the "daddies" of Pixel-J is Baidu, which has been engaged in the autonomous driving system for ten years.
"We have the same core part of AI algorithms as Baidu Apollo (Baidu's open platform for autonomous driving) and Radish Express (Baidu's driverless cab)."
Xia Yiping told CHNECAR that Pixel-J adopted Baidu's atomic core technical capabilities of voice, map, and smart driving, and did some secondary development and integration of their localities during development, thus creating a product experience more suitable for pushing to the C-end.
At the same time, Pixel-J has also built its own AI data iteration closed-loop system with the Baidu team, pushing the AI to learn and iterate on itself based on user usage behavior.
"So everything we have is reconfigured with a strong technical difficulty in it."
The algorithm system of intelligent voice and intelligent driving has now gradually left the era of hand-written code logic and entered the era of machine learning-driven by data. Mostly based on application scenarios, the engineers choose to build a suitable neural network after the input of massive amounts of data for training, so that it can learn to acquire image recognition, voice interaction, and even intelligent driving skills and eventually form an artificial intelligence that can perform certain tasks, that is, AI.
After completing the learning, AI will be quantified and "downscaled" into mature models with small data sizes to provide various services such as smart speakers, smart cars, and other end-side, and upgraded and iterated through new data learning.
The core of Baidu Apollo is a self-driving AI system that has learned to drive for ten years and has been quantitatively deployed in self-driving cabs like the Radish Express to achieve driverless capability in small areas with extremely high safety requirements.
Now, Pixel-J is bringing Apollo AI to be deployed in production vehicles to enable intelligent/autonomous driving capabilities in a wide area with the driver-monitored operation.
ROBO-01's 3D dynamic human-machine co-driving map interface
Of course, this process is not a simple "fetch", but requires a series of deep adaptations: for example, the lidar on the carrot fast-running vehicle is located on the roof, while the set of degrees for aesthetic and wind resistance coefficient considerations, the lidar layout on the front side of the vehicle. For this reason, the AI model needs to make adjustments to some data channels; for example, the adaptation and safety of the scenes of the Radish Express and the Pixel-J vehicle have different requirements, and the corresponding parameters and learning result requirements of the AI model will be very different.
But all these adaptations of secondary development and integration will not change the fact that Apollo AI already has the considerable driving ability. Just as a person who can drive a regular car can, after training and learning, get started in a race car.
While "experience" is one of the key reasons Tesla is ahead in the autonomous driving space, the nascent Pixel-J has more than a decade of experience integrated into it, and Apollo AI's experience is more urban-focused than other competitors' self-driving AI systems that run more at highway speeds. Baidu's self-driving cab fleet conducts real road tests in more than 30 cities across the country and has 27 million kilometers of safe, 0-accident self-driving test miles. These experience thresholds become the starting point for a set degree of intelligence.
Xia Yiping said that Apollo's more urban self-driving data is one of the important reasons why Pixel-J 's intelligent/autonomous driving capability performs better in urban scenarios and covers a wider range of cities at the beginning of landing.
"When building this autonomous driving product, we have to consider three parties at the same time: the user's demand for the product, the capability that the technology can achieve, and the restrictions of laws and regulations." Xia Yiping said to CHNECAR when explaining the technical path of Pixel-J in terms of intelligence, "Then we are actually more about how to maximize the intersection of these three."
03
Pulling up the 200,000 yuan level with Geely
In May, Baidu CEO Robin Li revealed that the first model of Guide is set to be a family car, priced at around 200,000 yuan. But whether it's LIDAR, NVIDIA's Orin X, Qualcomm's 8295, or a series of cool equipment such as 3D borderless integrated ultra-clear large screen, traceless side windows, various light sets, etc. are all very "costly", how to achieve cost control under the 200,000 yuan level?
"We hope to invest in intelligence to the best extent, other parts are more dependent on the cooperation between us and Geely, to get the space for cost optimization." Explaining to CHNECAR the cooperation model between both Pixel-J and its shareholders, Xia Yiping said.
"We go more in charge of thinking about how to build a smart product, spending time on product definition, applying the software and AI capabilities that Baidu has; combining Geely's hardware and engineering technologies with each other with the upper layer's need to achieve intelligence."
"Pixel-J and Geely are more in engineering, R&D, and manufacturing production related to this cooperation."
No screen blocking, foldable U-shaped steering wheel
The first model of Pixel-J has been confirmed to be produced in Geely's Hangzhou Bay factory. In addition, Xia Yiping said that the model of Pixel-J will also form a synergy with Geely in the supply chain. "We also share a lot of parts with Geely, so we still get a lot of support from Geely in terms of the whole cost control."
Pixel-J patent: jump light type, liftable and collapsible LIDAR located on the front hood
It is worth noting that not only the production and supply chain but also part of the intelligent technology capability of Pixel-J is based on Geely.
Achieving the high-level intelligent/autonomous driving capability of Pixel-J not only requires a mature core AI provided by Baidu, a high-calculus chip but also a strong enough set of electrical and electronic architecture to be able to support it.
At the press conference, Pixel-J unveiled its self-researched high-order autonomous driving intelligent architecture JET, which integrates electrical and electronic architecture EEA and SOA operating system to carry high-order autonomous driving core functions and AI capabilities to achieve full domain integration (inter-domain & intra-domain).
Among them, the Gigabit Ethernet backbone and SOA operating system with decoupled hardware and software are deeply connected with the high-bandwidth, highly decoupled open API ports of the vast architecture. It is also based on Geely's vast architecture that Pixel-J was able to pioneer the SIMUCar R&D process, starting intelligent development 15 months earlier than the industry average schedule, completing system security and stability verification, and delivering a full allocation.
In the fourth quarter of this year, Pixel-J will open its first brand-owned store in China and build its own sales, delivery, and after-sales system completely at the same time. As a latecomer, it will kill the market in most of the competitive years of 2023 with high-level intelligent driving capability. Whether Pixel-J, standing on the shoulders of Baidu and Geely, can truly become the "face" of automotive robotics, only the market will give an answer.
