EdgeAI
Book demo
All posts
August 14, 2026 · Andrew Mikhail

Matic proved people want to talk to their robots. The next frontier is where that voice runs.

Matic Cues brings voice and gesture control to the robot vacuum, with the wake word and the entire vision stack running on-device and command understanding going to Google's Gemini API in the cloud. Here's what that architecture says about the real state of on-device voice AI on a Jetson.

on-device-aivoice-airoboticsjetsonprivacy

On August 13, Matic shipped Matic Cues, and I think it's the most important consumer robotics release of the year.

Point at a spill. Say "Hey Matic, clean this." The robot hears you, sees you, locates the spot in 3D, and goes and cleans it. Say "follow me," and it follows. Say "vacuum the dining room rug and then mop the bedroom," and it sequences the job. It turns to face you when you speak. It hears you from ten meters away. It works in more than 70 languages.

And it shipped as a free over-the-air update to robots people already own. Nine years of work, handed to existing customers at no charge, in a year when the company's memory costs went up tenfold. That's rarer than it should be.

I posted about it that day and called it a masterclass in UX. I stand by every word. Matic understood something most robotics companies still haven't: the interface problem is the product problem. Co-founder Mehul Nariyawala said it better than any positioning doc I've read: "Humans have understood each other through cues for 200,000 years… the best way to talk to a robot is the same way we talk to each other."

I also wrote that the whole thing runs on-device. That part I got wrong. The correction turned out to be the most useful thing in the launch, because it points straight at where this industry's next hard problem lives.

Where Matic Cues runs, according to Matic

Not according to a reporter. From Matic's own product page, under a heading that reads "'Hey Matic' stays on Matic":

"The wake phrase is detected on-device. No audio leaves your home before the chime. Afterward, your request is processed anonymously by Gemini and never stored. Turn Cues off anytime in Settings."

Their engineering post on voice-data handling fills in the rest. Ambient audio is analyzed live by an on-device model and immediately discarded. Nothing is recorded, stored, or sent. After the wake word fires and the chime plays, the command goes to the Gemini API over an encrypted connection for anonymous parsing and multilingual understanding. Microphones ship off by default. You opt in, and you can toggle back out anytime.

The boundary is precise, and Matic drew it themselves:

StageWhere it runs
Ambient audio monitoringOn-device — analyzed live, immediately discarded
Wake phrase "Hey Matic"On-device
Gesture, pointing, 3D localization of the messOn-device — camera data processed locally, then discarded
Navigation, mapping, obstacle avoidanceOn-device
the chime— the line —
Natural-language command parsingCloud — Google Gemini API
70+ language understandingCloud — Google Gemini API

Bloomberg confirms it independently: "a cloud-based Google Gemini application programming interface, or API, is used to parse voice commands and to enable the comprehensive language support." Gizmodo notes it's the first Matic feature to require an internet connection.

To be clear about what this is: Matic disclosed all of it. On the product page, in an engineering blog post, and on the record with Bloomberg. Their wake-word detection really does run locally. Their mics really do ship off. Their camera stack never leaves the house. In a category where competitors route voice through Alexa and Google Home wholesale, this is the most privacy-respecting cloud voice implementation I've seen in consumer hardware.

So how did I get it wrong? I read the coverage, the coverage compressed a nuanced architecture into the phrase "on-device," and I repeated it to a few thousand people who work in robotics.

That points at a real problem, though not one of Matic's making. "On-device" now covers an enormous range of architectures. Full local inference sits at one end. Local wake-word gating in front of a cloud model sits at the other. Those are different products with different failure modes, and we have no vocabulary separating them. If someone who does this for a living can't tell from the coverage, buyers have no chance. We should fix that.

Why even Matic reached for the cloud

Matic is exactly the company that should have shipped fully local voice. Two founders out of the Nest camera team, who built the entire architecture around not sending your house to a server. In-house assembly in California. $115M raised, including a $60M round in January backed by Sutter Hill Ventures, the Collison brothers, Jack Dorsey, Daniel Gross, and Nat Friedman. Co-founder and CEO Navneet Dalal states the company's core tenet plainly: "families should not have to jeopardize their privacy to have their home cleaned."

They even had the silicon installed already: an NVIDIA Jetson Orin with 4GB of RAM.

That number is the whole story. Four gigabytes, already hosting a five-camera RGB-IR perception stack that turns the world into one-cubic-centimeter voxels and localizes in 3D while the machine drives around your home.

Now fit a conversational agent into what's left. Voice activity detection. Streaming speech recognition. A language model with tool calling. Conversation memory. Barge-in handling, so a user can interrupt mid-sentence. Text-to-speech. A real-time audio path. Across 70 languages. And none of it can take a cycle from the perception loop that keeps the robot from bumping a toddler.

You can't buy your way out with more memory either. Nariyawala, on the record: "Memory prices literally have gone from $12 for 4GB last year to about $120 now." He calls the 4GB ceiling a "good constraint" that forces the team to get everything they can out of the hardware they have.

He's right, and that constraint isn't going away for anyone. Every robotics BOM in 2026 is getting squeezed by the same memory market.

So Matic made the correct call for a company that needs to ship. Use Gemini, get 70+ languages and open-ended semantic parsing on day one, put the feature in customers' hands now instead of in eighteen months. I'd have made the same call. Shipping beats purity, and a voice feature people actually have beats a local model people are waiting for.

That's the real signal here. The strongest team in the category, with the right hardware and the deepest possible motivation, found this hard.

Why the last mile is still worth running

None of what follows is a knock on their decision. It's the honest engineering ledger of cloud voice on a robot, and the reason the local version stays worth building.

Latency becomes something you observe instead of control. On-device inference has a budget you can measure and hold. A network round trip has a distribution, and the tail is what people remember. I spent time at Humane watching what a multi-second gap does to someone's willingness to speak to a device again. They don't file a bug. They just stop talking to it.

Connectivity becomes a dependency where there wasn't one. Matic's robot previously worked over Bluetooth. Cues is the first feature to need the internet. Homes have dead zones, ISPs have bad nights, and the robot is in the room with you either way.

The privacy story picks up a seam. Not a breach, a seam. Matic's position was total and is now conditional, and the condition sits on the modality people feel most personally. Camera data is abstract to a customer. Words spoken in your kitchen are not. That's a communications burden a local stack never carries.

Inference becomes a recurring cost against a one-time sale. Hardware sold once, at a price already squeezed by tariffs and memory, with Cues given away free. Cloud inference scales with engagement. Nariyawala has noted that richer features, like a future sentry mode, would mean "greater reliance on the cloud." Local inference is capex you already paid.

Every one of these gets solved by running the stack on the machine.

Matic has already said what comes next

From Bloomberg, quoting the company:

"Once we get an understanding of the top commands used, we will distil a smaller on-device model for those who do not want to use Gemini API."

Nariyawala's target, per Gizmodo: get voice working "sans internet, at least within the 4GB RAM constraint," accepting it "wouldn't quite be as capable," with local-only customers still getting "80% of the benefit."

So the most sophisticated team in the category, having shipped the cloud version, is telling you the destination is on-device. They aren't defending the API call. They're treating it as a stage.

They're also being honest about the difficulty. Distilling a classifier for the top N commands is tractable. But Matic's own stated R&D direction is open-ended semantic understanding, interpreting something like "clean up the mess near the chair." A top-command distillation nails "vacuum the kitchen" and thins out on the long tail, which is where the delight of Cues actually lives. Getting a local model to hold both the common case and the long tail, inside 4GB, next to a live perception stack, is serious engineering. It's a research-grade problem wearing a product deadline.

That's not a criticism of their roadmap. It's the reason this layer deserves to be somebody's whole company.

The pattern isn't Matic's alone

Every robotics team arrives here in the same order. Years on perception, navigation, control, safety. Then someone asks how people will actually talk to the machine, and the honest answer is a second engineering org: voice activity detection, wake word, streaming ASR, interruption handling, small-model inference with tool calling, conversation memory, deterministic tool guardrails, text-to-speech, and a real-time audio runtime. Tuned for constrained edge hardware. Wired into your machine's APIs. Coexisting with safety-critical workloads without disturbing them.

No autonomy team wants that on their roadmap. No board wants to fund it. So it gets rented, which is a reasonable choice, and the one Matic made with better reasons than most.

But the components market doesn't close the gap. An ASR model is not a conversational robot. Assembling one from parts, fitting it into a Jetson that's already 90% committed, and connecting it safely to machine control is the actual work. It's also the work almost nobody wants to do twice.

What we're building at EdgeAI

EdgeAI is the on-device voice and agentic layer for physical AI. One C++ runtime, on the machine:

Speech → Reasoning → Guardrails → Robot Tools → Response

Local ASR and voice activity detection, interruption handling, small-language-model inference with tool calling, conversation memory, deterministic tool controls, and TTS. It's built to fit alongside an existing perception stack rather than compete with it, at a runtime footprint targeted under 10MB of RAM. Hardware-agnostic across NVIDIA, Qualcomm, Apple, and AMD silicon. EdgeAI is a member of the NVIDIA Inception program.

The guardrails matter as much as the models. The agent interprets intent and calls only approved functions on your machine. Navigation, obstacle avoidance, path planning, and safety decisions stay entirely inside your autonomy stack. We're the mouth and the ears, never the hands on the wheel.

I wrote C++ for machines that have to work in real time. Years on autonomous systems at NASA and Rivian, then voice at Humane. That's both halves of this problem: the machine that has to keep itself safe, and the person who won't wait more than a beat for an answer.

An honest ask

I'm not going to tell you what your constraints are. You know your Jetson's thermal headroom and your RAM budget better than I do, and after this week you should trust my read of a press release slightly less than you did.

But if you're shipping or piloting a robot on Jetson, Linux, or ROS, and voice is sitting on the roadmap under "figure out later," I'd like to have the conversation early. While it's still an architecture decision rather than a migration.

EdgeAI is opening a limited number of slots in our Design Partner Program. We'll evaluate a voice agent on your existing hardware, wire it to one real robot workflow, and tell you honestly if your compute budget can't take it.

No new box. No cloud. No touching your autonomy stack.

Get in touch, or email me directly at andrew@runedge.ai. Tell me what your robot needs to say.

To the Matic team: congratulations. Cues is the clearest demonstration yet that voice and gesture are how people will actually work with machines in their homes, and you shipped it to existing customers for free. You've made the case for this whole category better than those of us building the underlying layer have managed to. I hope you get it running entirely on that Orin. I think you will.

Sources

Primary — Matic's own materials

  • Hey Matic — Matic product page. "The wake phrase is detected on-device. No audio leaves your home before the chime. Afterward, your request is processed anonymously by Gemini and never stored."
  • How your voice data is handled — Matic engineering blog. On-device ambient analysis and discard; encrypted transmission to Gemini API after the chime; mics off by default.
  • Matic Cues announcement — press release. Dalal quote; on-device visual processing; FCC Covered List context.
  • Matic Innovation — five RGB-IR cameras, voxel mapping, local processing of camera feeds and maps.

Reporting

Back to all posts