The Tomy Omnibot, sold in 1984 for around $600 (roughly $1,800 today), promised a robot butler: remote control, cassette-tape message recording and playback, a serving tray. What it delivered was a motorized shell that moved however you pointed a remote at it — nothing resembling the autonomy the “butler” framing implied. Mario Cruz, a Miami-based maker, spent a restoration project closing that four-decade-old gap, published October 26 as “Omnibot Reborn: Giving a 1984 Robot Butler the Brain It Was Promised.”
The restoration itself wasn’t trivial. The original unit’s PCB had battery-acid damage to clean up, and its wheels disintegrated on the first test drive after sitting for decades. Cruz kept the original drive mechanism, though — and the original control scheme with it. The 1984 Omnibot steers by audio tones: 1614 Hz means forward, 2208 Hz means left, 1422 Hz toggles a relay routing audio to the onboard speaker. That system still runs the robot today; what changed is how the tones arrive. Instead of a cassette deck, a Bluetooth adaptor and speaker carry the same control tones wirelessly, synthesized by a Python script rather than played off tape.
The intelligence lives in a separate module, Ring-O, designed with Placitech (who handled soldering and enclosure design) and mounted on top of the Omnibot’s shell. It carries a Raspberry Pi 5 (8GB), a Sony IMX500 AI camera running YOLO11 nano at roughly 17ms inference across 80 COCO object classes, a 128×128 OLED functioning as the robot’s “eye,” plus its own cooling fan and battery.
The most instructive part of the write-up is what Cruz tried first and abandoned. The initial driving logic routed camera frames through an LLM — Groq-served Llama models — asking it to decide what to do next. That took about 15 seconds per decision and, per Cruz’s account, reliably returned “right” regardless of what the camera actually saw. He replaced it entirely with rule-based driving logic that decides in roughly 0 milliseconds, keeping the LLM only for a task it’s actually suited to: describing the scene out loud. It’s a clean demonstration of a pattern worth generalizing — an LLM is not automatically the right tool for a decision that needs to happen in real time, even when it’s technically capable of making the decision at all.
Getting YOLO’s bounding boxes to mean anything also took real debugging: fixing coordinate normalization, axis swapping, and letterbox padding to reconcile a 640×480 camera capture against a 640×640 model input — the unglamorous arithmetic that eats most of the time in any vision pipeline, long after the model itself is already working. The finished system includes a kids’ web interface at /kids, runs as a systemd service with auto-restart, and exposes a health endpoint that returns HTTP 503 when a subsystem fails rather than failing silently.
Full source is on GitHub under the MIT license. The robot has been shown publicly at Maker Faire Miami, where Cruz is a longtime organizer.