LiDAR Throttle and Stop Priority

This page records the current division of responsibility: the camera model owns steering, while LiDAR may cap forward throttle or request a hard stop inside the center safety corridor. LiDAR does not output steering.

Decision

lidar_avoidance.evaluate() measures the nearest valid point in the car-relative center corridor. runtime.py applies the result in both manual and autonomous forward driving when AEB is enabled:

  1. Normal. At or beyond 1.65 m, LiDAR does not reduce throttle.
  2. Progressive slowdown. Between 1.65 m and 1.25 m, the maximum target falls linearly from 100% to 60% reference throttle.
  3. Hold. Between 1.25 m and 1.05 m, the governor holds 60% reference throttle.
  4. Emergency stop. At or inside 1.05 m, it requests zero throttle and full braking (stop_reason = "lidar_emergency"). Reverse is excluded.
  5. Steering. Every steering command still comes from the driver or camera model. Points to the left and right are telemetry only.

The AEB controller toggle gates the entire LiDAR intervention policy. When it is off, occupancy can still be displayed, but the policy returns no throttle cap or stop request.

Alternatives Considered

Option Pros Cons
Fixed LiDAR left-or-right swerve Reacts without needing a vision label Can choose grass or another unsafe space because LiDAR does not understand sidewalk boundaries
Blend LiDAR clearance into model steering potentially smoother two steering authorities can fight; difficult to explain and validate
Model steering + LiDAR throttle/stop veto (chosen) one steering owner and explicit distance thresholds cannot steer around an obstacle by itself; stopping-distance and detection coverage still need measurement

Reason

Fixed swerve logic could steer away from an obstacle without knowing whether that side was still sidewalk. The current design keeps path choice with the vision model and uses LiDAR only for the narrower claim it can support: measured center-corridor clearance. The resulting throttle cap, emergency state, and stop reason remain visible in telemetry instead of being hidden in a blended steering command.

The current LiDAR is a Youyeetoo FHL-LD19 on USB (/dev/ttyUSB0 via a CP2102 UART-to-USB Adapter, 230400 baud). The reader auto-discovers the port and retries in a worker after disconnects, so serial reconnect work is not intentionally performed in the main control loop. Missing scans remain fail-open with respect to AEB.

How to Know It Worked (Test Gate)

  • Place a controlled obstacle at each threshold and verify normal, progressive slowdown, 60% reference hold, and emergency braking without a LiDAR-generated steering change.
  • Repeat in manual and autonomous forward modes with AEB on, then confirm AEB off leaves throttle untouched. Keep the car restrained during bench checks.
  • The stop reason, center clearance, occupancy, and lane-action telemetry record which branch fired.
  • Bench tools: code/test_files/lidar/lidar_uart_test.py and code/test_files/lidar/lidar_viewer.py.