More awareness: https://youtu.be/t4QeJrodqyU I'm building NeckGuardian -- a robotic arm that holds your phone at eye level and tracks your head to fix tech neck. One joint working. AI camera as the only feedback signal. 400ms lag. Servo at 60 deg/s. That's a 30-degree blind spot every command -- and why classical PID oscillates. In this video: the 400ms latency breakdown (ISP, ImageManip, YOLO, USB, Python), why the math is against bang-bang control, and the pulse-control state machine that actually converges. Plus the next blocker -- the YOLO model loses the arm at the setpoint because it was only trained on motion. Real numbers from real hardware: - Lag: 400 ms (OAK-D to Python) - Servo: 60 deg/s (MG90S continuous rotation) - Blind spot per command: 30 deg - Overshoot raw: 50 deg, after EMA(alpha=0.05) + 150 deg/s clamp: 12 deg - Pulse formula: pulse_time = error / 60 deg/s; cooldown 500 ms - Detection rate moving: 95 percent, at setpoint: 50 percent - YOLO fine-tune dataset: 50 images, 6 keypoints Long arc: NeckGuardian : then MPC controller, then vision-language-action (PhyAI), then full multi-joint. Pausing in August. Chapters 0:00 Hook - I thought my PID gains were wrong 0:19 A quick question 0:39 Can you guess, here is the answer 1:08 PID gains can’t fix it 1:23 Answer was in the problem only. 1:35 Control theory: pulsed approach 1:49 Pulse control state machine 1:55 The next blocker: AI blind at setpoint (50 percent) 2:10 Roadmap: ZeroWhip, NeckGuardian + micro-car, MPC, PhyAI 2:13 Pause + how to follow Portfolio YouTube - @PhyAIEngineer -- PhyAIRobotics - @QubitGarage -- Quantum Computing IG - @technology.spiritual -- all-niche hub - @tech.biohacking -- sensor-validated - @rahuldeo2047 -- real me Code: github.com/rahuldeo2047 #robotics #physicalAI #controltheory #MPC #YOLO #OAKD #NeckGuardian #ZeroWhip #buildinpublic #computervision #edgeAI #DIYrobotics #tendondriven #PETG #ESP8266 #servocontrol