Drone pursuit terminal built with Grok Bot marks where tracking loses confidence

BuBBliK @k1rallikOriginally posted 2026.10.27

Open original on X

I built a drone pursuit terminal with Grok Bot this f*cking thing follows him off the roof. I wanted to see what happened when someone actually tried to break the track. Sharp turns. Dropping to another level. Disappearing behind concrete. So I gave Grok Bot the camera footage and one instruction: “Keep track of this person. When you lose him, show me exactly where you stopped being sure.” here’s the build step by step: step 1 → load the drone video into a local app. Everything stays on the computer. step 2 → detect the person, draw the box, assign one track ID. That ID follows the current track through the scene. step 3 → add motion prediction. He goes behind a wall → LOCKED becomes COASTING → the tracker looks for him near the predicted position. step 4 → if the prediction stops being reliable, clear the box and switch to SEARCHING. Pick the person up again when there’s enough visual evidence. step 5 → reconstruct the visible surroundings from the camera frames. Roof edges, walls, ground, trees. Keep the right panel aligned with what the camera actually sees. step 6 → put the same target box on that panel. When he drops to the lower level, both views update together. step 7 → project the estimated depth into a local overhead view. Now you can follow his position relative to the camera and surrounding surfaces. step 8 → graph image movement, relative target size and tracking confidence underneath. Every jump and disappearance leaves a trace. step 9 → export the whole thing as one synchronized video. Grok Bot writes the pipeline. The local vision models handle the frames. Watch the section where he drops into the building and runs behind the concrete. That was the part I kept replaying. Full build prompt below ↓

This post quotes another post: x.com/k1rallik/status/2114726480559112502