Matic is a pretty amazing robot. It also happens to clean your house. Point at a coffee spill on the kitchen floor, say “Hey Matic, clean this,” and the robot turns to look at you – physically pivoting toward your voice, the way a pet or person might — and sees what you’re pointing at, motors over, and cleans.
You don’t need an app, a pre-mapped home space, or pre-labeled rooms.
That’s impressive.
But it also, despite the company calling it the “most American” robot available, doesn’t yet pass recent FCC rules about foreign robots.
Matic Cues is the update the Mountain View company pushed out last week to every Matic already in a customer’s home. The software upgrade makes it the first home robot that understands more than 70 spoken languages and universal human gestures, the company says.
Matic carries five cameras and a four-microphone listening array. The wake word – think “Hey Siri” or “Alexa” – runs entirely on the device, so ambient room audio never leaves the robot. Because it has a mic array, Matic can locate which direction your voice came from. Proprietary on-device software then finds you, sees your hand and your finger and even your eyes, then triangulates the target you’re pointing to against its own position in your house. Camera data never leaves the machine, and the microphones ship off by default, the company says.
Of course, on-device compute for a little robot vacuum doesn’t actually speak 70 languages: the actual command you say after the wake word is anonymized and sent to Google’s Gemini AI engine.
On July 28, the FCC added foreign-produced “advanced robotic devices” to its Covered List, tech it says poses an unacceptable national security risk. The definition is wide enough to include most robot vacuums and robot lawn mowers: over 4.4 pounds, wireless connectivity above 200 kbps, environmental sensors, autonomous navigation. Devices that have already been authorized do keep their authorization and can keep selling, though the Commission layered on restrictions around firmware updates.
New models face a much harder path in, but there’s a domestic-end-product exemption at 65% U.S. content by value, rising to 70% in 2029.
Matic designs and assembles its products in California and markets itself as the most American robot in its category, but it does not currently clear that 65% bar.
We’re the most American robot in the market, but we still have work to do in meeting the guidelines set by USG.
co-founder Mehul Nariyawala told me via email. And, on whether the FCC action is a windfall:
Candidly, we have no idea how the FCC move will play out.
John Koetsier: How does Matic understand pointing?
Nariyawala: Matic has 5 RGB cameras and 4 Microphones. When users say “hey Matic”, it uses 4 microphone array to determine the direction of the audio and turns towards the users. This is critical as when we call another human or even pet, they look at us. This is the same. Matic looks at the users using its cameras. First it finds users, then it looks for their hand, then the finger, then eyes, and then it triangulates what users are pointing towards with its own position. It’s a proprietary algorithm that we’ve trained that works entirely on device.
John Koetsier: Are all 70+ languages processed on-device?
Nariyawala: Yes. “Hey Matic” is all on device so nothing is going to cloud in terms of ambient sounds. The gestures are all detected on the device as well. No camera data ever leaves your robot. Once you say, hey matic, and you hear a chime, you can give a command. This command uses Gemini API to parse it anonymously. No user info is sent. We chose to do it this way because it enables 70+ languages and many commands. Once we get an understanding of the top commands used, we will distill a smaller on-device model for those who do not want to use Gemini API. In the meantime, if you don’t want to use Matic Cues, you can turn it off in the settings. And microphones will remain turned off.
John Koetsier: What LLM are you using? On-board? What kind of CPU/GPU drives that?
Nariyawala: For “hey Matic” we’ve built our own model. To parse the commands in different languages, we use Gemini API.
John Koetsier: Is this a step toward a general-purpose home robot?
Nariyawala: It’s a step towards Rosie the Robot. We want to put Rosie the Robot/Alfred from Batman in every home. However, our approach is different. Most of what is being built in robotics today is technology looking for a problem to solve. Humanoids. Foundation models. World models. They are extraordinary feats of engineering but they are not products. Families do not want robots. They want self-cleaning homes. They want organized homes. They want to stop thinking about chores completely.
Growing Rosie the Robot: First perception — teaching Matic to see and move through the world. Then manipulation. Then long-horizon planning. We are growing Matic through exactly this sequence.
We’re the most American robot in the market, but we still have work to do.
Read the full article on Forbes.