Every autonomous vehicle runs the same three-stage loop: sense the world with lidar
or cameras, plan a safe path, then actuate steering/throttle to follow it — repeated
tens of times per second. This top-down 2D view casts the same fan of rays as the
3D version, straight down onto the track plane.
rays[k] = cast(pos, heading + k*Δangle, senseRange)
avoidance = Σ (1/dist_k) * perp(ray_k) for obstacles seen
lateralOffset += (avoidance*gain - lateralOffset*0.5) * dt
steer = pathTangent + lateralOffset
- Vehicle speed — forward speed along the planned route.
- Sensor (lidar) range — how far ahead the fan of virtual lidar rays can detect obstacles.
- Steering gain — how aggressively the controller reacts to a detected obstacle; too low causes collisions, too high causes overshoot.
- Obstacle density — how many random obstacles populate the track.
This sense → plan → act cycle, with lidar-style perception and a reactive local
planner, is the same architecture used by real self-driving cars, warehouse robots
and planetary rovers — just running far more sophisticated perception and planning
stacks.