AI Needs a Body: Three Builds from Tomorrow Lab® Where AI and Hardware Work as One

A drawbridge watcher, a pet food bowl that knows who's eating, and a connected leak-detection system show what happens when intelligence is designed into the object from the beginning.

Most people first met AI through a text box as they asked a question and got an answer that appeared superior to previous methods, responding with more depth and accuracy. It hasn’t taken long for AI to evolve from search-engine replacement to wide-spread software integration.

Although it seems AI has exhausted its reach through subliminal spaces, possibly a more interesting shift happening now is that AI is moving into physical objects, allowing that object to see, decide and act in the real world, often without a screen anywhere in sight.

But the process of getting here takes more than just adding an AI model to a finished product. An AI system is only as capable as the sensors that feed it, the mechanism it controls, and the power budget it runs on. Decisions like these are left to hardware experts that can consider what the intelligence can actually do. When the virtual and physical are designed together from the very beginning of a project, the opportunities become endless.

Tomorrow Lab®, a 16 year old hardware invention studio in Brooklyn, has lately been exploring the different ways AI and hardware can integrate, from local computer vision to tiny on-chip neural networks to fleets of connected devices that behave like never before.

AI that sees

Not far from the studio sits a drawbridge that opens on its own schedule. The team set out to build a system that watches and learns the bridge schedule to know, alert, and predict others when it is up, so nobody gets ‘bridged’ and has to wait.

The build uses a webcam and the Arduino UNO Q, a board that pairs a Qualcomm Dragonwing QRB2210 processor running Debian Linux with an STM32U585 microcontroller for real-time control. The Linux handles a webcam feed and an AI vision application through Arduino App Lab, while the microcontroller properties stays open to the hardware considerations that Arduino has always been known for.

The result is a small device that turns a live camera feed into a simple, useful answer about the state of the bridge. A few years ago, a system like this would have meant streaming video to a server. Today, the system can run offline inside of a box the size of a sandwich, which opens the possibilities of adding a visual perspective to almost any object that needs to understand what is physically in front of it.

AI that decides

Tomorrow Lab also applied the AI<>hardware crossover in a way that pet parents could benefit from. One pet finishes its own food and then moves on to the other pet’s food bowl. The team designed a feeder that recognizes which pet is eating and only serves the one it belongs to.

The heart of this build is the Analog Devices MAX78000, a microcontroller with a dedicated convolutional neural network accelerator built into the chip alongside an Arm Cortex-M4 core and a RISC-V coprocessor designed to run image recognition at very low power, offline, at the edge. Instead of sending images to the cloud, the bowl makes its decision on the device in real time.

Low-power inference means intelligence can live inside small, everyday objects that would never justify a full computer. Recognition happens instantly with no ongoing cloud cost and the images never leave the bowl. The feeder is a playful example, but the same approach applies to access control, quality inspection and any product that needs to tell one thing from another.

AI that connects

The Tomorrow Lab team developed a suite of safety devices that can detect sources of leaks to reveal problems before they turn into real damage, using Golioth, a cloud platform built for managing connected hardware.

A single gas or water sensor can tell you that something is leaking. A network of sensors reporting to a shared platform can do far more. With device management, data streaming and over-the-air updates handled through Golioth, sensor readings from every device can be gathered in one place, analyzed for patterns and acted upon. If a certain sensor’s readings are higher than others, it could reveal the true source of a growing leak. The devices can also improve after they ship, since firmware and detection logic can be updated remotely, without a trip back to the manufacturer.

Here, AI and hardware reinforce each other through the hardware that gathers data from the physical world, while intelligence turns a stream of readings into a decision a homeowner or facility manager can actually use.

What the three builds share

Each project started with a real-world problem rather than a pure desire to add AI. A drawbridge that blocks your route, pets stealing each others’ dinner and a mysterious leak hiding behind a wall are all physical problems that software alone cannot solve.

Each also paired the job with the right silicon or platform. High res vision with a full operating system went to the UNO Q, low-power recognition went to the MAX78000 and fleet-scale sensor intelligence went to Golioth. That matching is a design skill in its own right, and it is where most of the invention happens.

The next wave of AI products will look beyond apps and more into physical objects, from home safety and health monitoring to mobility and children's products. The teams like Tomorrow Lab® that build them will need to think in atoms and algorithms at the same time, choosing sensors, chips and mechanisms with the same care they give to models. The three builds featured on Potentially Genius offer a working preview of what that looks like.