A hobbyist is building his home AI "a body", one sensor at a time

A developer has spent two months documenting, in a personal essay on "Towards AI", how he's giving his home AI system a physical body. Not all at once, however, but sensor by sensor.

The system (which he calls Hana) started with pan-tilt-zoom cameras and has since learned a detailed spatial map of the house. It corrects itself in real time until it could locate a red sofa in the dark, unprompted. This week, according to the author, he added a Viatom VTM-20F fingertip pulse oximeter over Bluetooth, which gives the system read-only access to a person's blood-oxygen and pulse readings — with a hard rule that it must report "no data" rather than fabricate a number when no finger is present.

The next planned upgrade is a relay-controlled garden irrigation valve, and this is where the author says he's deliberately slowing down. Any autonomous watering command would need a hardware-level shutoff timer sitting outside the AI's own reasoning loop, so a software error can't leave a valve open indefinitely. A similar relay on the property's gate has been ruled out entirely for now, since a stuck-open valve can flood a garden while a stuck-open gate is a security failure.

For object detection, the setup reportedly uses a fast YOLO11 model for rough person-detection alongside a slower vision-language artificial intelligence model for identification, after a smaller model reportedly hallucinated a person out of a coat rack on a porch camera.

The author basically describes this project as proof of the fact that embodiment doesn't really need be a humanoid, sentient robot. One just needs an accumulating set of sensors and very deliberate, strict limits on what each one is allowed to control.