What can Jidu compare with?

Author: Shiyun Zhu

Editor: Lingfang Wang

GEE·denka’s mass-produced cars finally see the light of day, proving that they are not just another “cartoon car maker” and attempting to answer the question that has been chasing them for 605 days: How do they compare to top players like Xiaoli Wei, Xin Wang, AITO, and Zeekrbi when it comes to being “late?”

On October 27th, GEE·denka officially released its first automobile robot, ROBO-01 Lunar Limited Edition, with a price tag of CNY 3.998 million. Although it is a limited edition with only 2,000 cars, its hardware components of the intelligent driving system, electronic and electrical architecture, tri-electric system, and a wind resistance coefficient of 0.249cd will be consistent with the subsequent mass-produced cars.

At the Guangzhou Auto Show in November, GEE·denka will release a large batch of mass-produced ROBO-01 and launch them in the third and fourth quarters of 2023 alongside the Lunar Limited Edition. When it comes to pricing, GEE·denka CEO Xia Yiping told Electric Vehicle Observer that it is very competitive.

The information provided by the Lunar Limited Edition allows us to catch a glimpse of GEE·denka’s core competitive strategy as a “latecomer”:

“The first wave of electric vehicles has already pulled the product foundation of electrification relatively flat, and the products tend to be homogeneous. But the competition for intelligence is just beginning, and this is the field where GEE·denka has a core competitive advantage. We need to do accurate positioning, achieve something that others lack or we excel in, and focus on the core product concept and experience in the field of intelligence. We hope to use technology to change this industry,” Xia Yiping said. “If the customer experiences something great, they will definitely pay for it. When we sell ten thousand cars a month, it’s when I can finally breathe a sigh of relief.”

Have GEE·denka honed their products to a sharper point in the already “rolled-up” field of intelligence?

How to build an automobile robotThrough mass-produced products, we can better understand Jidu’s concept of “automotive robots” with three abilities: natural communication, free movement, and self-growth. The entire technology (design and system) of Jidu is moving towards this direction. Xia Yiping used ROBO-01’s cockpit as an example to explain the implementation process of Jidu’s concept into products.

ROBO-01 lunar version is equipped with Qualcomm’s fourth-generation 5nm Snapdragon digital cockpit platform – the 8295 chip, which has a computing power of 30 TOPS, nearly 8 times the computing power of the 8155 chip. It also has a 35.6-inch 6K integrated screen, with a 95% NTSC high color gamut and a 10000:1 ultra-high contrast.

These hardware components support a 3D intelligent driving map that can react to real-time road conditions and a speech interaction system that deploys a complete speech AI model on the client-side. The full client-side voice recognition speed is as fast as 0.5 seconds, and the end-to-end response speed is within 0.7 seconds, completely breaking away from the dependence on network signals.

This means that after the passenger gives a voice command in the car, the interaction system “understands” the command within 0.5 seconds, and it takes less than 0.7 seconds to execute the command. In other words, the voice system can complete the passenger’s voice command within 1.2 seconds, reducing a lot of time for “thinking” and “reacting”.

Behind this is Jidu’s support for the “natural communication” ability of automotive robots.”The core requirement of ‘natural communication’ is to understand the speaker’s true intention based on the context and semantic meaning, rather than rely on specific ‘keywords’ to trigger the ‘command set’ in the system to execute commands. However, the accuracy of natural language neural networks (AI) depends on the size of its models.
Larger models can calculate higher precision parameters, thereby obtaining more accurate answers. However, the deployment of large models in end-to-end devices such as cars is often limited by computing power, and pruning is required to reduce calculation precision to enable the vehicle to operate smoothly.

To achieve the goal of ‘natural communication’, Jidu has selected the high-performance Qualcomm 8295 chip with higher process and computing power to deploy complete models. “We are not stacking components just for the sake of doing so, but instead configuring the corresponding hardware based on the targeted functional goals,” says Xia Yiping.

The idea of configuring hardware based on functional requirements is also reflected in Jidu’s intelligent driving technology package. The ROBO-01 lunar exploration edition’s intelligent driving hardware includes dual NVIDIA Orin X with 508 TOPS, two Heysai AI128 LiDARs with a detection distance of up to 200 meters, seven 8 million pixel cameras, four 3 million pixel cameras for side view, and one 2 million pixel camera, as well as five millimeter-wave radars and 12 ultrasonic radars.

ROBO-01 lunar exploration edition can achieve point-to-point intelligent driving capabilities, with the ability to smoothly connect different road conditions from high-speed to city and parking domains.

In terms of perception solutions, Xia Yiping revealed that Jidu will initially adopt a perception fusion scheme that integrates heavy sensing and high-precision maps.”In August, the Office of the Ministry of Natural Resources issued a notice announcing the pilot application of high-precision maps for intelligent connected vehicles in six cities including Beijing, Shanghai, Guangzhou, Shenzhen, Hangzhou, and Chongqing. In late October, licenses for the urban advanced driver assistance maps were issued in Guangzhou and Shenzhen, with Baidu being the first batch approved to support the intelligent vehicle with city navigation and driving assistance features. As a joint venture of Baidu, Jidu’s intelligent driver capabilities will also be the first to be implemented in these six cities.

Although Jidu’s intelligent driving hardware will use high-precision maps in the early stages, the hardware is also quite “heavy”. It is currently unique in the industry with its dual lidar and 7 8-megapixel cameras configuration, which will be fully equipped on Jidu’s mass-produced vehicles.

Why is the perception hardware so “heavy”?

According to Xia Yiping’s interview with “Electric Car Observer”, the choice of hardware configuration is also based on the needs of intelligent driving functions. “It is also a compromise. We can do some functions without a lidar, but the ability may be much smaller. For example, in low-speed or parking scenarios, we have more advantages with the lidar.”

“But this does not mean that we think the higher the performance of the hardware, the better. We always pursue the most suitable algorithm configuration and the most efficient hardware. In the future, after the evolution of algorithms, we may reduce from 8 million (pixels) to 5 million, or even 3 million. But at this stage, we hope that the whole system can be a little safer, so some aspects are slightly redundant.”

This redundancy is not only embodied in the perception hardware. In addition to using dual Orin redundancy, Jidu’s mass-produced vehicles also have a double redundancy safety function of emergency side parking through the 8295 in the cabin.

It’s worth noting that Jidu also has a high reuse of Baidu’s autonomous driving neural network. Based on the L4 level Robo-Rabbit, Jidu was able to run its high-speed and urban domain fusion navigation and driving capability in nine months. At present, it can not only make unprotected left turns, but also perform U-turns and turnarounds.

The “Extreme” Behind the Vehicle Type

In fact, in order to achieve differentiation from similar products on the market, Jidu made its first product somewhat “extreme”.

ROBO-01 Lunar Edition has no hidden door handles, but instead completely removes the handles. The four electric front doors and four millimeter-wave radars can not only automatically open when the car key is near, but also be opened by voice; Inside the car, the U-shaped steering wheel removes the shift lever, and most of the central control buttons are cancelled.

In terms of vision and experience, the lack of physical buttons may be very “cool”, but it also extremely tests the software capabilities behind the product, whether it can achieve high usability and reliability to avoid inconvenience and even risks caused by the lack of physical buttons in some extreme situations.

In response to this, Xia Yiping told Electric Vehicle Observer that Jidu has achieved reliable functional guarantees for “extreme” vehicle models through its understanding of software and JET electronic and electrical architecture.Currently, Jet has over 2000 members, among which nearly 70% are software engineers. They have developed Jet’s SOA operating system, which has achieved the industry’s first integration of cabin and chassis through SOA. Moreover, it only takes 1.5 hours to load the software onto the production end, which enables the vehicle to function normally.

“We have already prepared for cross-domain integration at the software layer, and we only need to wait for a system-level chip (that can support it) to achieve the hardware-level domain fusion. And 1.5 hours (loading speed) is extremely impressive in the industry,” said Xia Yiping.

Under the SOA operating system, Jet’s EE architecture includes four domain controls: intelligent driving, intelligent cabin, whole vehicle control, and powertrain. They are connected by a high-bandwidth gigabit Ethernet loop, linking the entire vehicle’s sensors and high-performance computing platforms.

“First, it is about understanding the software, and then being able to truly engineer these functional understandings. At the same time, it is important to fully understand and predict the future directions of chips, sensors, and intelligent technology. This is a systematic ability, and Jet’s core in ensuring the reliability of ‘extreme’ vehicle models,” Xia explained.

Jet has completed all of next year’s parts and capacity reservations, and all links in production and sales coordination have been established. “Now we need to spend more time building the service and sales system to ensure the brand’s user experience. The battle between Jet and me is far from over, and there will be more challenges ahead,” Xia concluded.

This article is a translation by ChatGPT of a Chinese report from 42HOW. If you have any questions about it, please email bd@42how.com.