My Robot's Eyes Lag 400 ms — Why PID Fails and the Pulse Fix That Doesn't

I'm building NeckGuardian, a vision-guided robotic neck (Phase 2a). This week I hit a wall that every controls engineer eventually meets: the eye-to-actuator reaction window. The loop sees a target, but there's a ~400 ms lag between detection and motion. Classic PID makes it WORSE here. It overshoots, oscillates, and the head hunts around the target instead of settling. The fix isn't tuning gains harder. It's changing how you think about the command: stop sending continuous corrections, start sending discrete PULSES with a cooldown longer than the reaction window. 15 deg steps, 60 deg/s, 0.25 s move, 0.5 s cooldown beats the 0.4 s lag, so it converges. Steady-state error drops to about 2.1 deg and it HOLDS. Then the twist: at rest, the system's AI presence-detection confidence drops from 95 percent to 50 percent. Solving the motion problem opened a new blind spot. This is real hardware from my bench, not a simulation explainer. Chapters: 00:00 The reaction window 00:13 The 400 ms eye lag 00:21 Why PID fails here 00:38 The rig 00:44 The pulse fix 00:56 control_loop.log 01:09 AI detection drops at rest 01:25 The new blind spot 01:30 What's next Rahul (@PhyAIEngineer) — Physical-AI / robotics engineer. Next video: does fine-tuning fix the blind eye? #robotics #controlsystems #PID #physicalAI #embeddedsystems #roboticsengineering