KI for Robotics

  • home_icon HOME.
  • Research
  • Research Institutes
  • KI for Robotics
  • Highlights

Total 29

Development EV-Based Urban Self-Driving Algorithm

How high is the current level of self-driving. Self-driving is realized in a highly stable manner in situations where the driving conditions are relatively simple and the road features are clear, such as high ways and urban express ways. However, the story is totally different in complex urban areas. There are so many things that should be taken into consideration, including flashing traffic signals, traffic jam on the road, cutting-in vehicles and pedestrians, and unclear lanes markings. Even Tesla, which is leading the self-driving technology, is unable to implement the autopilot function in all these situations.
The research group in the KAIST Institute for Robotics, led by Professor Hyunchul Shim, is broadening the scope of self-driving complicated urban areas. The researchers demonstrated their competence by winning the championship in the Hyundai Motors Autonomous Driving Challenge held in November 2021 and receiving an award from the mayor of Seoul. They also published two SCI papers by presenting and an adaptive cruise control methodology that is applicable to complex traffic situations.

|For the Future of Korean Self-Driving System!|

Urban self-driving vehicle without GNSS
In preparation for the competition, the research group developed several algorithms to deploy Level 4 autonomous vehicles in the complex urban environment (self-driving on roads with possible unexpected incidents in complicated urban areas, alleys, and curved roads). One of the key contributions was to deal with the complicated urban environment as an HD-map(3D point cloud map + 2D semantic road map) so that vehicle can estimate its position in the world accurately. The high-accuracy localization algorithm based on HD-map provided less than 10 cm positioning error without GNSS in the high-rise buildings areas, like the Sangam-dong course for the competition. Moreover, Multi-modal sensor fusion and deep learning-based perception stacks enabled to detect the dynamic obstacles and the complex situations. Subsequently, the motion planning and decision-tree-based behavior planning algorithm were in charge of collision avoidance of nearby vehicles. In the intersection scenarios, V2X-communication signals (traffic regulation) were transit to the V2X router, and vehicles could consider the traffic signal during passing the intersection areas. Among these algorithms, the construction of the HD-map has contributed as a great step forward to the victory of the competition. The research group said it was a huge difference from other team; the KI team was the only one without GPS among the six teams that proceeded to the final, and the team became the champion of the challenge.
“Of course, self-driving vehicles equipped with expensive GPS systems and sensors show better performance, but only when GPS is nat interfer. Their challengeable scenario is the environmental features are unclear such as missing lane markers, blinded objects. That’s why deploying autonomous vehicles in urban areas is still difficult. We think that the HD-map-based autonomous system can serve as a complementary product before fulfilling high-quality deep-learning-based autonomous system, which can replace a human being..” says Daegyu Lee (Doctorate student in Department of Electrical Engineering at KAIST), the team head in the KAIST Institute for Robotics.

Opening a new horizon from a niche
Director Shim thinks that the urban self-driving technology level of Korea is not so high now. The urban environment, which we humans are also confused about, requires of AI human intelligence level for clear understanding. However, the AI level is not so high. The leaders of self-driving, such as Tesla, are moving in that direction, but the research conducted in Korea is far behind them in terms of the input resources. However, Director Shim is looking for through which the research group can open a new horizon.
“We cannot compete with the leaders in terms of quantity. Nevertheless, there are areas that are unseen to the leading groups of self-driving. If we continue to take efforts in such areas, we will be able to pioneer a new area, which may appropriately be called ‘Korean self-driving technology.’

Prof.Hyunchul Shim
2021 Annual Report

KAIST 291 Daehak-ro, Yuseong-gu, Daejeon (34141)
T : +82-42-350-2381~2384
F : +82-42-350-2080
Copyright (C) 2015. KAIST Institute