A software developer is gradually building a physical presence for his home artificial intelligence system, not through a humanoid robot, but by connecting a growing network of sensors, each bound by strict operational limits. In a detailed personal essay, he describes a two-month experiment to give what he calls "embodied" awareness to an AI assistant named Hana, while deliberately restricting its ability to act.
Building spatial memory through vision
The system initially gained awareness of its environment using pan-tilt-zoom cameras, which allowed it to construct and continuously refine a detailed spatial map of the house. The developer reports that this real-time self-correction reached the point where Hana could independently locate a red sofa in complete darkness, without being prompted to search for it.
Adding biometric sensing with strict boundaries
This week, the setup expanded to include a Viatom VTM-20F fingertip pulse oximeter connected over Bluetooth, giving the AI read-only access to blood-oxygen saturation and pulse rate. A hard-coded rule forces the system to return "no data" whenever a finger is not detected, ensuring it never invents or guesses a health metric. The developer frames this constraint as a non-negotiable safety layer, preventing the AI from fabricating physiological readings.
Controlling physical infrastructure safely
The next planned addition is a relay-controlled garden irrigation valve, a step the author says he is approaching with deliberate caution. He intends to install a hardware-level shutoff timer that operates completely outside the AI’s reasoning loop, so a software error cannot hold the valve open indefinitely. By comparison, a similar relay for the property’s gate has been ruled out for now, given the different risk profile: a stuck valve might flood a garden, but a stuck-open gate represents a clear security breach.
For visual detection, the developer uses a fast YOLO11 model to roughly identify a person’s presence, then a slower vision-language model handles identification. This two-stage approach replaced an earlier, smaller model that reportedly hallucinated the presence of a person from a coat rack on a porch camera.
The project, as described, serves as a working argument that machine embodiment does not require a sentient humanoid form. Instead, it can emerge from an accumulating set of sensors, each paired with carefully defined, strict limits on what it is permitted to control.
Sources: pub.towardsai.net, unsplash.com, unsplash.com