Cat robot hack · 29 Sep 2026

The screen works. The escape doesn't yet.

The demo screen and the memory already look like the pitch; the robot's escape is mediocre, and letting the cloud planner steer made it worse. Everything below was recorded tonight against the fake Elegoo in the simulated room.

Needs you · one decision

Should the cloud planner steer the robot, or only narrate on screen?

Narrate only: the planner's mode and reason show on the dashboard and in Weave, and the reflex drives. The demo still shows every sponsor
Steer, but only escape_corner, and only when no cat is in view
Steer fully, as now, and tune the planner prompt on Wednesday

Evidence is one 60 s run each, so noisy: 45% of the time caught when the planner's modes were being overwritten by a bug, 74% once they reached the wheels. Its "patrol" and "escape_corner" pull the robot off running away.

Three views · what a judge would see

1The recorded run shows what the robot sees beside where it actually is.Left, the camera with the reflex's reason; right, the room from above: blue robot, red cat.74% caught

54 s, fake Elegoo on seed 6 with the cloud supervisor on, recorded 29 Sep 00:18. Watch it scan, spot the cat, spin away and run; then lose it near the bottom wall, where the cat pins it and fills the frame with legs. The orange mode label is the cloud planner's. main.py --dash --demo-video against fake_elegoo.py, composited with ffmpeg.

2The live screen tells the story a judge needs in one glance.Cat boxed with bearing and closeness, the reason for each move, the planner's mode, the vision verdict.0.93 conf

dashboard: cat boxed at 0.93, mode flee, vision verdict — localhost:8000 on the simulator, 28 Sep: YOLO 0.93, planner "flee" in 0.51 s, vision verdict in 0.82 s, loop 112 fps. The layout is unchanged since; this frame predates tonight's fixes.

3The memory answers in English and hands back the clips, which is the video-agent pitch.Asked through the dashboard after the recorded run.12 clips

dashboard memory answer with three playable clips — "When did the cat get too close, and did I get away?" Answer: closest approach 73% of frame height at 08:18:12, and the robot escaped. Clips are playable. The captions are wordy and describe one middle frame, so they sometimes miss a cat seen elsewhere in the clip. Playwright screenshot, 29 Sep.

What recording found · all but the last fixed tonight

·The cloud call froze the robot about half the time; it now runs in the background.fixed

Commands per minute went from 477 to about 2,200; watchdog stops from 21 to 1.

·The loop overwrote the planner's mode every frame, so its decisions never reached the wheels.fixed

One line in main.py; the orange label now changes.

·The close-cat rule (distance sensor under 25 cm with a recent sighting) is in, as the default.fixed

It fired 74 times in one run: "cat too close to see: back away". Change NEAR_CM in reflex.py if you want another option.

·Clips were in a codec browsers cannot play; the dashboard now converts them.fixed

H.264 on first request, cached.

·Captioning ran at 17 s per clip tonight, and the dashboard captions at question time, so a question on stage could hang for minutes.open

Fix for Wednesday: caption in a background thread as each clip closes, so questions only search.

Repo github.com/robertnowell/catbot at cce7128, 14 tests passing. The simulator's cat is relentless and slow; a real one sprints, pauses and loses interest, so treat these rates as behaviour checks, not predictions.