Urban Autonomous Racing

The race

The 2022 Segye AI Robot Car Race, hosted by Segye Ilbo at Sunmoon University, was the first autonomous race in Korea with a simultaneous start. The format was 10 laps of a 1.2 km track, just over 12 km, within a one-hour limit. Organizers could drop obstacles on the track to force lane changes, and every car had to pass an emergency-stop test from at least 10 km/h. A single-lap time trial set the grid, and then all seven cars started the final together. Finishing without a collision was the main goal.

System

Layer	Components
Platform	ERP42 UGV with its internal motor controller
Compute	Neousys POC-545, NVIDIA Jetson Xavier
Sensors	Velodyne VLP-16 LiDAR, RGB camera, u-blox GPS, low-cost IMU
Software	Ubuntu 18.04, ROS Melodic, OpenCV 3.2, PCL, YOLOv5
Simulation and debugging	MORAI simulator, RViz

The stack has three jobs: track, detect and avoid. Path tracking follows a pre-recorded GPS race line (RDDF), steering against lateral error with a look-ahead distance tuned to balance cornering against high-speed stability.

We originally planned a full camera-LiDAR fusion pipeline. Getting NVIDIA drivers, CUDA, OpenCV and ROS to agree on versions became a major bottleneck, so we cut scope to a LiDAR-first system and used the competition partly to collect data.

Perception

LiDAR. Avoidance runs on the point cloud. We split the 360° scan into three zones (ahead, adjacent left and adjacent right), filtered each by distance threshold, and reduced them to a boolean occupancy per lane. Close-range emergency stops rely on LiDAR alone.

Vision. The camera model identifies obstacles 10–15 m ahead and within about 30° of the heading, early enough to plan a lane change. No race-specific dataset existed, so we built one. From 56,000+ frames of video recorded at a previous competition, we sampled every 7th frame to get about 8,000 candidates. We added cases where an obstacle sits on the right during a left turn, which had caused trouble on track. We then labeled about 4,000 images in Roboflow and split them into train, validation and test sets for YOLOv5.

Detector output (ERP42 boxes) on onboard footage from the 2022 race. Video recorded by the UOS Robotics team.

The detector worked well in normal driving, but had clear failure modes. It sometimes detected a vehicle’s right wheel as a second, smaller object, and its boxes often clipped the wheels. It also missed vehicles at very close range and handled overlapping vehicles poorly. Neither case was in the training data. We enabled YOLOv5’s class-agnostic NMS and tuned multi_label to suppress duplicate boxes. The remaining failures need data: more close-range and overlap examples, then a higher confidence threshold.

Control

Longitudinal. Relative velocity from the sensors was too noisy to use for car-following. We commanded speed from distance alone: Target Velocity = k * (d - s)^n, where d is the distance to the obstacle ahead and k, s and n are tuning parameters. This gave smooth, predictable deceleration behind other cars.

Lateral. The first lane-change logic used timers and flags, and it oscillated. The car would start a new avoidance while already in one and wiggle across the lane. We replaced it with a three-state machine. From DEFAULT the car enters CHANGING_LANE and commits to the maneuver until it finishes. It then enters COOLDOWN to settle before it may consider another change. This removed the oscillation.

Results

Tracking, localization and detection held up in the race. Avoidance at speed did not. The car detected obstacles but started lane changes too late, so it often reached its emergency-stop distance before the maneuver finished and came to a full stop instead of passing. That latency kept us from completing the race cleanly.

Takeaways

• Multi-stage maneuvers need explicit states; timers and flags were not enough.
• Building control on the most stable measurement (distance) paid off over a richer but noisy one.
• Data is the lever for perception. Next is an active-learning loop that flags close and overlapping cases for labeling.
• The real bottleneck is decision latency: avoidance must trigger earlier, possibly with fused sensor cues for the state transitions.