RoboRacer & Autonomous Driving ์ฐ๊ตฌ ํ
๋ง
1. ์ํ ์์จ์ฃผํ ํ๋ซํผ(RoboRacer, 1/10-scale) ๊ธฐ๋ฐ ์๊ณ ๋ฆฌ์ฆ ์ฐ๊ตฌ
RoboRacer(F1TENTH) ์ฐจ๋์ ํ์ฉํ์ฌ ์ค์ธ๊ณ ์์จ์ฃผํ ์๊ณ ๋ฆฌ์ฆ์ ์คํ ๊ฐ๋ฅํ ํํ๋ก ์ถ์ยท๊ตฌํํ๋ ์ฐ๊ตฌ๋ฅผ ์ํํฉ๋๋ค.
- SLAM, Localization, Mapping
- Obstacle avoidance(Follow-The-Gap, AEB), Pure Pursuit ๊ธฐ๋ฐ ๊ฒฝ๋ก ์ถ์ข
- ์ต์ ๊ฒฝ๋ก ์์ฑ(Optimal Trajectory)๊ณผ ๊ณ ์ ๋ ์ด์ค ์ฃผํ
2. Embodied AIโWorld Model + ๋ชจ๋ธ์ฒดํน ๊ธฐ๋ฐ ์๋ฎฌ๋ ์ด์
์์ ์ฑยท์ ๋ขฐ์ฑ ์ฐ๊ตฌ
์์จ์ฃผํ ์์ด์ ํธ๊ฐ ํ๊ฒฝ๊ณผ ์ํธ์์ฉํ๋ฉฐ ํ์ตยท์ถ๋ก ํ๋ Embodied AI ํ๊ฒฝ์์,
World Model๊ณผ Formal Verification(๋ชจ๋ธ์ฒดํน)์ ๊ฒฐํฉํ์ฌ ์์ ์ ์ด๊ณ ์ ๋ขฐํ ์ ์๋ ์๋ฎฌ๋ ์ด์
๋ฐ ์ฃผํ ์ ์ด๋ฅผ ์ฐ๊ตฌํฉ๋๋ค.
- World Model์ ํตํ ํ๊ฒฝ ๋์ญํ ์์ธก ๋ฐ ๋ฏธ๋ ์ํ ์๋ฎฌ๋ ์ด์
- Timed Automata ๊ธฐ๋ฐ ํ๊ฒฝยท์ํฉ ๋ชจ๋ธ์ ์์ ์ฑ ๊ฒ์ฆ
- ๋ชจ๋ธ์ฒดํน์ผ๋ก ๋์ถ๋ ์ ์ฝ(safety constraints)์ RLยทplanning์ ์ฐ๋
- Sim-to-real ์ ํ ์ ๋ฐ์ํ ์ ์๋ ์ํ ์ํฉ ์ต์ํ
- ์๋ฎฌ๋ ์ดํฐ ๋ด ์๋๋ฆฌ์ค ์์ฑ์ ์์ ์ฑยท์ผ๊ด์ฑ ๊ฐํ
ํต์ฌ ๋ชฉํ:
"ํ์ต ๋ฐ ์ ์ด ์๊ณ ๋ฆฌ์ฆ์ด ์์ ์ ์ฝ์ ์๋ฐํ์ง ์๋๋ก ๋ณด์ฅํ๋ ๊ฒ์ฆ ๊ธฐ๋ฐ(verification-aware) Embodied AI ์์จ์ฃผํ ํ๋ ์์ํฌ ๊ฐ๋ฐ."
3. ์๋ฎฌ๋ ์ดํฐ ๊ธฐ๋ฐ ์์จ์ฃผํ ํ์ต ๋ฐ ํ๊ฐ
์ค์ฐจ ์คํ์ ์ํ์ฑ๊ณผ ๋น์ฉ์ ์ค์ด๊ธฐ ์ํด Gazebo / Isaac / custom simulator ํ๊ฒฝ์์ ์๊ณ ๋ฆฌ์ฆ์ ๋ฐ๋ณต ๊ฒ์ฆํฉ๋๋ค.
- RL ๋ฐ imitation learning ์ ์ฑ
ํ์ต
- ๋์ ์ ์๋๋ฆฌ์ค ์๋ ์์ฑ ๋ฐ stress-testing
- World Model ๊ธฐ๋ฐ predictive simulation๊ณผ ๊ฒฐํฉํ ์ฐ์์ ํ์ต
- sim-to-real ์ฑ๋ฅ ๊ฒฉ์ฐจ ์ต์ํ ์ค๊ณ
4. Formal Verification์ ์ ์ฉํ ์์ ์์จ์ฃผํ
์์จ์ฃผํ ์๊ณ ๋ฆฌ์ฆ์ ์์ ์ฑยท์ ๋ขฐ์ฑ ๋ณด์ฅ์ ์ ํ ๊ธฐ๋ฒ์ผ๋ก ์ฒด๊ณํํฉ๋๋ค.
- Timed Automata ๊ธฐ๋ฐ ์ฃผํยท์ถฉ๋ยท์ ๋ ์๋๋ฆฌ์ค ๋ชจ๋ธ๋ง
- ์ถฉ๋/์ ๋ ์กฐ๊ฑด์ ๋ํ safety property ๋ชจ๋ธ์ฒดํน
- Verification-guided control ๋ฐ runtime shielding ์ค๊ณ
- ์์ ๋ฏธ์ค์ ์ํฉ์ ๋ํ ์๋ ๋ฐฉ์ด๋์ ์ ์ด
5. ๊ฒฝ๋ AI / ์ธ๊ณ ๋ชจ๋ธ(World Models)์ ์์จ์ฃผํ ์ ์ฉ
LLMยทWorld Model ๊ธฐ์ ์ ์์จ์ฃผํ ํ๊ฒฝ์ ๋ง๊ฒ ๊ฒฝ๋ํํ๊ณ ์ค์๊ฐ์ฑ ์๊ตฌ์ ๋ถํฉํ๋๋ก ๊ตฌํํฉ๋๋ค.
- World Model ๊ธฐ๋ฐ ํ๊ฒฝ ์์ธก ๋ฐ ๋ฏธ๋ trajectory ์ํ๋ง
- ๋ถํ์ค์ฑ์ ๊ณ ๋ คํ planning ๋ฐ ์ ์ด
- Multi-modal state representation ํ์ต
6. ๊ต์กยท์คํ์์ค ๊ธฐ๋ฐ ์์จ์ฃผํ ์ํ๊ณ ๊ตฌ์ถ
RoboRacer Korea๋ฅผ ์ค์ฌ์ผ๋ก ๊ต์กโ์ฐ๊ตฌโ๊ฒฝ์ง๋ํ ์ํ๊ณ๋ฅผ ๊ตฌ์ถํฉ๋๋ค.
- ROS2 ๊ธฐ๋ฐ ๋ฉ ๊ณผ์ ๋ฐ ์ค์ต ์๋ฃ(Lab-06~Lab-12) ๊ฐ๋ฐ
- ์ด๊ธ~๊ณ ๊ธ ์ํฌ์ ๋ฐ ์ค์ฐจ ์ค์ต ๊ธฐ๋ฐ ๊ต์ก ์ปค๋ฆฌํ๋ผ ์ด์
- ์ ๊ตญ RoboRacer ๋คํธ์ํฌ ๋ฐ ์คํ์์ค ํ๋ก์ ํธ ํ์ฑํ
RoboRacer & Autonomous Driving Research Themes
1. Algorithm Research Based on Small-Scale Autonomous Driving Platform (RoboRacer, 1/10-scale)
We conduct research to scale down and implement real-world autonomous driving algorithms in an experimentally feasible form using RoboRacer (F1TENTH) vehicles.
- SLAM, Localization, Mapping
- Obstacle avoidance (Follow-The-Gap, AEB), Pure Pursuit-based path following
- Optimal trajectory generation and high-speed racing
2. Embodied AIโWorld Model + Model Checking-based Simulation Safety & Reliability Research
In an Embodied AI environment where autonomous agents interact with the environment to learn and reason,
we combine World Model and Formal Verification (Model Checking) to research stable and reliable simulation and driving control.
- Environment dynamics prediction and future state simulation through World Model
- Safety verification of environment and situation models based on Timed Automata
- Integration of safety constraints derived from model checking into RL and planning
- Minimization of risk situations that may occur during sim-to-real transfer
- Enhancement of stability and consistency in scenario generation within simulators
Core Objective:
"Development of a verification-aware Embodied AI autonomous driving framework that ensures learning and control algorithms do not violate safety constraints."
3. Simulator-based Autonomous Driving Learning and Evaluation
To reduce the risks and costs of real vehicle experiments, we repeatedly validate algorithms in Gazebo / Isaac / custom simulator environments.
- RL and imitation learning policy training
- Automatic generation of challenging scenarios and stress-testing
- Continuous learning combined with World Model-based predictive simulation
- Design to minimize sim-to-real performance gap
4. Safe Autonomous Driving with Formal Verification
We systematize safety and reliability assurance of autonomous driving algorithms through formal methods.
- Modeling of driving, collision, and braking scenarios based on Timed Automata
- Model checking of safety properties for collision/braking conditions
- Design of verification-guided control and runtime shielding
- Automatic defensive action control for safety non-compliance situations
5. Application of Lightweight AI / World Models to Autonomous Driving
We lightweight LLM and World Model technologies and implement them to meet real-time requirements for autonomous driving environments.
- Environment prediction and future trajectory sampling based on World Model
- Planning and control considering uncertainty
- Multi-modal state representation learning
6. Building an Education and Open-Source-based Autonomous Driving Ecosystem
We build an educationโresearchโcompetition ecosystem centered around RoboRacer Korea.
- Development of ROS2-based lab assignments and practice materials (Lab-06~Lab-12)
- Operation of beginner to advanced workshops and real vehicle practice-based educational curriculum
- Activation of nationwide RoboRacer network and open-source projects