highway-sim-eval · Interactive explainers

Step-by-step animations behind the experiments in highway-sim-eval. Every number on these pages comes from a real run in the repository. Pages are in Chinese; each card has an English summary.

Big picture

MapSimulation testing as an exam智驾仿真全景地图Scenario library (exam questions), simulator (exam room), system under test (examinee), evaluation (grading), and the data loop around them; open vs closed loop, MIL to road, ISO 26262 vs SOTIF, and where this project sits. Ten short lessons, one question each. LayersNo simulator is complete: five kinds of exam room仿真的五种考场Engine simulation (CARLA, MetaDrive), log replay, neural reconstruction (NeRF / 3DGS), world models, and HIL / road / shadow mode, ordered from cheap and high-volume to expensive and rare. Each one animated: what is real, what is fake (replayed cars drive through you; reconstruction blurs off the recorded path; a world model makes a car vanish), what it tests, and how real-car failures loop back as new tests.

Control and planning

A3Pure Pursuit look-ahead前视距离试驾场Short look-ahead weaves, long look-ahead cuts corners: 1.23 / 0.27 / 0 m overshoot at 2 / 5 / 15 m. A7Lane change, slowed downS 形变道慢放A quintic (minimum-jerk) lane change vs a straight diagonal at 10 m/s: steering wheel, lateral acceleration and d / speed / acceleration curves in sync. Left, centre, right, centre; half the time means 4x the lateral acceleration. A7Why plan in Frenet: a lane change in a curve弯道变道两种画法Same left curve, same Pure Pursuit. Offsetting the path toward the road's left (Frenet) ends in the left lane; adding d to map y ends back in the own lane (3.50 m off) once the road points up. Lane change plus curve peaks at 3.19 m/s². A3 · metricLateral error, done right离中心线多远Distance to the nearest waypoint hides up to 0.25 m of error; point-to-segment distance fixes it. A8Controller lag changes safety metrics刹车慢半拍Same planner, lagging PID: minimum gap drops from 4.96 m to 4.36 m. A8Car following in a curve弯道跟车：两种开法IDM + Pure Pursuit + bicycle model, with the lead car braking inside the curve. A8Stop-and-hold with hysteresis停车保持的门槛Different enter and exit thresholds stop the creep-stop-creep cycle at low speed.

LiDAR perception

A5One sweep of a 32-beam LiDAR激光雷达扫一圈Rings on the ground, occlusion behind the lead car, the blind circle around the ego car. A5Where LiDARs are mounted激光雷达装在哪Roof, roof-front tower, bumper, headlights, full surround: a draggable 3D comparison. A5DBSCAN step by stepDBSCAN 怎么分堆Core points, expansion and noise; why eps must sit between the largest gap and the nearest neighbour. A5Why one car became two clusters车顶上的缝A 0.86 m unscanned gap between the hood face and the roof splits the car at eps 0.5. A5From points to a box一堆点怎么画框Axis-aligned min / max boxes, scanned one axis at a time. A5Why the box center is biased前车框为什么偏I-shape vs L-shape visibility: the center always shifts toward the sensor. A5Only the rear face is visible前车只剩车尾The lead car box is 2.30 m short, and no eps setting changes that. A5Ground removal on a slope坡道上去地面On a 4% upgrade a flat height cut leaks 3,371 road points; a RANSAC ground plane leaks 0. 3D steps plus a side-view slow motion.

Tracking

A6Kalman filter: predict, then measure卡尔曼：猜一步，量一下One frame of a 1D Kalman filter on a distance number line: predict, measure, blend by confidence. On a frame where differencing reads −7.5 m/s, the filter only nudges its speed from −1.81 to −1.97. A6Multi-object tracking: who is who多车跟踪：谁是谁Three cars, radar boxes with no names in shuffled order. Each frame: predict every track, match by a distance table and the Hungarian algorithm with a 2 m gate, then manage IDs. Includes a 5-frame occlusion (keep 10 frames: same ID; keep 3: the car comes back as a new ID) and a guardrail false alarm that lives 1.1 s as a ghost track.

Evaluation

A5Matching detections to ground truth框认领真东西Greedy one-to-one matching on center distance; recall 0.40 / 0.60 / 1.00 at nuScenes gates 0.5 / 2 / 4 m.

Simulation platform

B1Same driver, two simulators: move first or slow down first先挪还是先减速Lead car brakes to a stop: the update order puts it at 9.5 m (A4) vs 10.5 m (HighwayEnv), true 10.0 m; my hardest brake −3.69 vs −3.44.