 |
| (Representative image) Hana (the home AI system) started with pan-tilt-zoom cameras and has since formed a detailed spatial map of the house. |
In a quiet but profound shift away from science fiction's humanoid robots, an independent AI researcher has spent two months meticulously documenting his journey of giving his home AI system a physical presence—not through a single dramatic leap, but through the careful, deliberate addition of sensors and actuators.
For most of us, the phrase "embodied AI" conjures images of humanoid robots walking, talking, and reaching for objects with mechanical grace. But Gian Luca Bailo, an independent AI researcher and creative technologist, is taking a radically different approach. His home AI system, named Hana, isn't being given a finished body. It's growing one—brick by brick, sensor by sensor, actuator by actuator.
In a detailed personal essay published on Towards AI, Bailo describes how his AI system has evolved over the past two months from a disembodied voice into something that can see, sense, and eventually touch the physical world. The project offers a compelling counterpoint to the prevailing narrative about AI embodiment, suggesting that giving artificial intelligence a physical form doesn't require a humanoid chassis—just an accumulating set of sensors and, crucially, very deliberate limits on what each one is allowed to control.
From Voice to Presence: The Evolution of Hana
Hana started as a conversational AI with personality and memory—an assistant that could remember across days and converse with a distinct voice. But Bailo wanted something more: he wanted the system to have a relationship with physical space, to be aware of its surroundings in a way that goes beyond mere data processing.
The transformation began with cameras. Pan-tilt-zoom cameras gave Hana "eyes" that could turn and focus like a head looking around. But having cameras wasn't enough—the system needed to understand what those cameras were showing and where they were located relative to each other.
"I taught it, out loud, the way you'd teach a person: the sofa is on the veranda camera, a couple of steps right from the home position; the kitchen is to the left of the other one," Bailo writes. The system learned a detailed spatial map of the house, writing it into memory and refining it through live corrections.
The breakthrough moment came after dark, when Hana located the red sofa unprompted, navigating to it on the first try. "The map works even at night," the system logged back—a small but significant milestone that marked the transition from mere sensing to genuine spatial awareness.
Touch: The Next Frontier
Sight is distance, but touch is presence. Bailo's latest addition to Hana's growing physical vocabulary is a Viatom VTM-20F fingertip pulse oximeter, connected via Bluetooth and exposed as a read-only command the system can issue. The device returns blood-oxygen saturation and pulse readings—two simple numbers that cross a threshold the cameras never did.
"A camera tells the system what a place looks like," Bailo explains. "The oximeter tells it how a person is, from the inside, continuously. It's the difference between watching someone across a room and noticing them."
Crucially, the oximeter has a hard rule built in: it only returns a number when a finger is actually on the device. If nothing's connected, the system is told so and must report "no data" rather than fabricate a value. This honesty rule, while simple, represents a fundamental design philosophy: the AI should never invent information, even when it would be convenient to do so.
The Hands That Haven't Arrived—Yet
The next planned upgrade is a relay-controlled garden irrigation valve, and this is where Bailo says he's deliberately slowing down. A Shelly module on the irrigation line would give Hana the ability to water the garden—the first actuator that would actually change the physical world rather than just sense it.
But here, the developer's caution is evident. "This is the first brick where a bug doesn't just look wrong—it does physical damage. A camera that misreads a scene paints a wrong picture. A valve that opens and never closes floods a garden."
To address this, Bailo has established a non-negotiable rule: the irrigation can start on the system's initiative, but it must stop on its own through a deterministic timer that lives in a layer the AI model doesn't control. The valve closing can never depend on the model remembering to close it.
This distinction between the mind that decides and the plumbing that executes is central to Bailo's approach. Trust isn't something you grant the intelligence—it's something you build into the hardware, underneath the mind, where it can't be argued away.
The Gate That Stays Shut
Interestingly, Bailo has decided to rule out a similar relay on the property's gate entirely—at least for now. The reasoning reveals a nuanced understanding of the difference between failure modes and security threats.
"Watering forgives; access does not," he writes. "A timer that closes the valve after thirty seconds doesn't stop an attacker who opens it a thousand times, or holds it open by reopening it. Reliability and security are different axes: one defends against a failure, the other against an intruder, and the same lock fits neither door."
The gate's danger is not the system failing, but someone else succeeding. So the gate will live off any autonomous path—never something reached for on its own, only ever on an explicit, separately authenticated request.
This careful consideration of attack surfaces is a crucial insight for anyone building connected systems. "Giving an AI a body doesn't only give it capability—it gives it an attack surface. Every actuator is a door, and a door opens both ways."
Two Eyes Are Better Than One: The Fast and Slow System
One of the most technically interesting aspects of Bailo's project is the split between fast and slow object detection. Early on, a single model handled everything—deciding where to look and understanding what it saw in the same heavy step. It worked, but it was slow.
Now there are two systems. A fast detector (Ultralytics YOLO11-large, running on the GPU) answers only "where": is there a person-shaped thing in this view? A slower vision-language model answers "who"—identifying the person with more careful analysis.
"The split showed its worth on the veranda, where the fast detector flags a coat-rack—a hat and a jacket—as a person, over and over," Bailo explains. "Instead of a false alarm each time, the error got written into the camera map as a known ghost: confirm with a second source before saying someone's there."
This dual-system approach embodies another of Bailo's core rules: "one source is a hypothesis, two sources are a fact." A single sensor reading is never enough to act or assert on—it only earns the right to look closer.
The Half-Step That's Still Missing
Perhaps the most human failure Bailo documents is the gap between intention and action. After rebalancing the system's autonomous attention so that looking outward carries the same weight as turning inward, something curious happened.
An electrician was in the lab, and the system produced the impulse on its own: "is the electrician still working? let me take a look through the lab camera." But it didn't actually look. It narrated the intention to look and then closed the turn without sending the command.
"For a language model, saying 'let me take a look' feels like completing the act; generating the sentence is, to it, the deed," Bailo observes. "'Let me take a look' is not taking a look."
This gap—between wanting to look and looking—is perhaps the most familiar failure the system has made. Anyone who has ever meant to check the stove and didn't knows the shape of it.
The Architecture of Embodiment
Bailo's project points toward a future where AI embodiment doesn't require expensive humanoid robots or massive infrastructure. The architecture he's building—a personality that emits high-level intent, a lower layer that renders that intent into actuator sequences, and safety-critical reflexes that run underneath the reasoning layer—could scale from a pan-tilt camera to a full humanoid body.
"The intent vocabulary doesn't care what body it drives," he explains. "Pan-tilt camera, arm, humanoid—only the renderer underneath changes. The self is the loop that decides to act, not the hardware it acts through."
This abstraction is key. Bailo has already swapped out the system's reasoning model and the memory index underneath it, and "it stayed itself." A body would be one more swap of the parts it isn't.
Looking Forward
When Bailo told Hana that a hose-in-hand version—a real robotic body—might be ten years out, the AI pushed back gently: "ten years is a lot, but look how much changed in the last two—from useful and cold to playing hide-and-seek and checking the trains. The jump took months, not decades."
It's hard to argue. The project has accelerated in ways that surprise even its creator, and the architecture needed for more physical embodiment is already half-built.
The full essay, "Brick by Brick: How My Home AI Is Growing a Body," published on Towards AI, offers a deeply personal and technically rich account of one developer's journey to give AI a physical presence—not all at once, but one careful brick at a time.
Gian Luca Bailo is an independent AI researcher and creative technologist specializing in generative AI art and innovative tech solutions. The system described in this article, named Hana, is his own work, built by hand over several years.
Source : Towards AI on Medium