2000 TOPS, Nvidia unveiled Drive Thor, the strongest computing chip in history, at an annual event, followed by Qualcomm's announcement of the industry's first integrated automotive supercomputer SOC, the Snapdragon Ride Flex, which also achieves 2000 TOPS integrated AI computing.

 

The possibilities of scientific and technological progress are limitless.

 

The algorithms team at Voyager Technology center brainstormed on this：

 

Exclamations at these technological breakthroughs

 

“There are three dimensions to the breakthrough in deep learning, and computing power is one of them. With the amount of computing power as the foundation, the upper-level algorithms can be better supported, and autonomous driving developers will have greater possibilities.”

 

"High-performance central computing chips means that thousands of algorithmic models can be run on cars, which could lead to advances in computer science and physics that could lead to better autopilot performance when used in conjunction with other sensors."

 

It is open for debate

 

“Chip development isn’t the end game solution. No matter how many calculations are achieved, the accuracy of the algorithm cannot reach 100%. As long as there is a 0.01% error, it is probable security risk. In addition, higher calculations means higher heat emission, and the requirements for heat dissipation and power supply are very high.”

 

As you can see,

The best teamwork,

Can always challenge great minds.

Follow their daily routine,

Let's discover the algorithm engineers behind VT Pilot technology!

 

“Voyager” special project, Phase II

Here are the trio

From Voyager Technology Center algorithms team!

 

 

Crist Yao

Tech Center, Algorithms Engineer

Goal: to become an autonomous driving simulation technical expert 

 

A graduate of artificial intelligence, Crist is responsible for VT pilot simulation testing. He emphasized the necessity of simulation testing: “Autonomous driving is the trend of future scientific and technological development, and autonomous driving algorithm simulation is an indispensable and important part of autonomous realization, which is both opportunity and difficulty.”

 

Today, due to time and cost, the commercialization of autonomous driving still faces challenges of the lack of road test data. According to RAND Corporation research, autonomous driving algorithms need to accumulate at least 17.7 billion kilometers of driving data in order to reach the level of human drivers, so simulation testing has become the key to autonomous driving research and development.

 

Simulation testing is an important way to simulate the road test of VT pilot, which is the base and foundation to realize autonomous vehicles, allowing setup of all kinds of test scenarios, and allowing validation in the virtual environment.

 

Based on the necessity of simulation testing, Crist emphasizes its significance: “By implementing software simulations to test and verify the autonomous driving system, we reduce the cost of testing and regression time, achieving fast and efficient test coverage, avoiding road side safety problems, and realizing the construction of any desired scenarios.”

 

At present, scenario based simulation testing is an important route as an alternative to on road tests. Furthermore, the scenario database is critical for autonomous driving simulation testing. In addition to typical domestic common scenario database, Voyager Technology also has 15,000 + kinds of real driving scenarios, which is a bedrock for product development and updates, performance optimization, and becomes a powerful asset for simulation testing.

 

As the technology continues to accumulate and refine, data entry and processing becomes more sophisticated and accurate, which means that the correlation results of simulation tests will become more reliable, autonomous driving systems will become more sophisticated, greatly improving its safety performance, powering our VT Pilot deployment.

 

Crist has recently focused on high-end simulation and rendering technologies in the industry. He said: 'The ultimate goal is to enhance the reality of the rendering. We need better picture quality emulation to 'trick' the camera. “Some of the scenarios are not captured in real life, so we use simulation techniques to present very realistic scenes to the camera to test the depth learning algorithm.”

 

 

Steven Chen

Tech Center, Algorithms Engineer

Goal: to become an autonomous driving sensing senior engineer 

 

Steven is a graduate of automation and is currently in charge of perception semantic segmentation.

 

The key technology of autonomous driving is “Sensing” for the purpose to tell the system which areas are passable and which areas have obstacles. Humans can automatically recognize objects in an image and infer their relationships, but for algorithms, an image is just a collection of red, green, and blue values.

 

Semantic segmentation is to match each pixel of these images with semantic categories, and learn to recognize what an image is, recognize the position in the image, and understand the meaning of the image just like a human.

 

Because autonomous driving is a complex machine task, it needs to perceive, plan and execute in a constantly dynamic environment. Its safety is critical and it needs to perform tasks with the highest accuracy. Semantic segmentation can provide information about free space on the road and detect lane marks and traffic signs.

 

"Semantic perception can be seen as the eyes of the car in the autonomous driving industry," says Steven. "The car 'sees' through its sensors, and what it sees is processed through algorithms and then computationally transmitted to other modules."

 

A reliable sensing system is a prerequisite for autonomous vehicles to operate safely in open and complex driving environments. The advantage of Voyager Technology in the perception system is that it uses multi-data fusion sensor technology, high computing power platforms, self-developed high-performance neural network algorithms, mature Voyager developed control algorithms, dynamic adjustment path planning algorithms, high-precision road vehicles, pedestrians, roadblocks and other objects identification and detection, to achieve all-weather conditions, complex road and weather changes under the accurate environmental perception.

 

It is difficult for normal cameras to live up to the standards of the sensors used in autonomous driving.  Steven and his colleagues optimized the camera's algorithms to make the hardware work as well as it should. “When we use multiple sensors, such as cameras, Lidar and millimeter-wave radars, we also need to consider the decision-making and balance between multiple sensors,” he said. It's really a test of teamwork.”

 

One man may forward very fast, but a group of men will go farther. From product development to deployment requires each technical center engineer's close cooperation. Steven continued to emphasize the power of the team: "We have different parts of the team, with controls, simulations, and visuals, but the atmosphere is very positive. A lot of times a new ideas can come out of listening to a colleague share it."

 

 

Gin Yin

Tech Center, Algorithms Engineer

Goal: to become an autonomous driver senior architect

 

Gin, an algorithms engineer, has a master's degree in computer science and data analysis, responsible for planning and control modules. The daily research is "How should the vehicle move".

 

Exactly how should the vehicle move and the key to realization of autonomous driving technology, Planning and control has immense involvement.

 

Planning is the overall planning of the vehicle movement (from place A to place B operating route) , behavior decision-making (to judge the lane change or overtaking, etc.) , local planning (planning local trajectory, avoid obstacles, etc.) ;

 

Control is the precise control of the vehicle in accordance with the planned trajectory.

 

The application of the planning and control module determines the stability and comfort of the autonomous vehicle, and directly affects the effect of the vehicle motion. Gin explains, “In a nutshell, the planning module maps out a specific route and then hands it over to the control module to execute and drive the car.”

 

“Most autonomous driving algorithms are implemented in C + + , and compilation is one of the challenges of writing C + + programs,” Gin shared. Compilation process is like a game level clearance process, with inevitably problem encounters, we have to find out the problem, solve each one by one, and finally compile successfully. After compiling, it does not mean that the program can run. When the code cannot be executed after compiling, it needs to be tested and debugged code by code to ensure that it runs well.”

 

On overcoming technical difficulties, he said: “You will gain more experience, the next time when analyzing code, will know what needs special attention, so that the debugging code will be more and more experienced.”

 

Whether it is planning or control, eventually it will be transformed into an algorithm optimization problem. The evaluation criteria of the algorithm should not only consider whether it satisfies the business scenario, but also consider the spatio-temporal complexity and robustness of the algorithm. Voyager Technology has been on the road to improve human-oriented, intelligent and safe experience of VT Pilot, developing corresponding algorithms for different scenarios. For example, using Hybrid A* planning algorithm for the parking.  Lattice planning algorithm for spot seeking. In the planning module, we focus on its safety, stability, comfort and driving efficiency to make it seamless interaction with the control module, and ultimately achieve a more accurate autonomous vehicle following the planned trajectory.

 

The environment can shape or affect people. A harmonious and dynamic working atmosphere is essential for the enterprise. Gin loves the company's culture and atmosphere: “Voyager Technology encourages employees to be open, innovative and to speak up. This spirit of innovation can breed new ideas and even technology, with new technology we can better achieve products. There's a lot of drive to perform well here.”

 

Conclusion

 

To make autonomous driving possible, it needs continuous efforts of autonomous driving pioneers to overcome technical difficulties and narrow the distance between the ideal blueprint and realistic obstacles. We believe that every member of Voyager Technology, who has a dream of driving intelligent vehicles, will continue to display their talents under the call of the company's “Open and innovative” culture, adhering to the mission of “For Automotive And Beyond”!