Here is the failure signature that should scare you: the cloud broke when the robot got good.
Not when the model was hallucinating. Not when the policy was jerky and knocking over cups. It broke at the moment of success — when the robot became reliable enough that everything else in the room became the weakest link.
That inversion is the whole story, and most of the industry hasn't lived it yet.
Start with the lab, because the lab is where the wrong intuition gets manufactured. In the lab, cloud inference feels obvious: more compute, easier debugging, the identical setup your research team already runs. Cameras grab frames, ship them over Wi-Fi to a rack of GPUs in a data center, the model decides, the answer comes back. Strong network, predictable latency, effectively unlimited compute. Everything works.
Then Sunday Robotics put the robot into real houses. And the architecture that felt obvious started dying in strangers' living rooms.
The bugs weren't in the model. They were in the building. Dead zones in bedrooms. Neighbor interference bleeding across walls. Misconfigured mesh routers doing something creative with your packets. And the kill shot most people never check: do you know your 300Mbps Xfinity may only have 5Mbps upload?
That asymmetry is the quiet assassin. A robot streaming camera feeds out to the cloud lives and dies on upload bandwidth, and the number on the box — the number the customer paid for — describes download. The uplink is a tenth of it, or worse, and nobody told you.
It reportedly took their engineers hours per home to debug. Per home. When the evaluation crew came back, reportedly the person who built WhatsApp video call looked at it and said, "Yep, that's what you get in the wild."
Sit with that. The person who spent a career making real-time media survive hostile consumer networks looked at the failure and recognized it instantly. This is not an exotic edge case. This is the baseline condition of the American home, and the robotics industry priced it in at zero.
Here is why it stayed invisible so long, and why it's a trap. The problem only became legible once three things were true at the same time.
One: the robot was fully autonomous. Two: it was being evaluated in real, unseen homes, not the lab. Three — the crucial one — the policy was reliable enough that network jitter became the dominant failure mode.
Before all three are true, the network is never your top bug. A shaky policy hides the network. Lab Wi-Fi hides the network. A tele-operated demo hides the network. You have to remove every other cause of failure before this one surfaces at the top of the stack.
As the team put it: before that, you do not even get to work on the real blockers to mainstream market adoption.
That sentence is the load-bearing insight of this entire piece. The blockers to mass adoption are gated. You cannot see them from where most teams are standing, because most teams haven't shipped a good enough policy into enough unseen homes to make the network the tallest problem in the room.
The blockers to mass adoption are gated. You cannot see them from where most teams are standing, because most teams haven't shipped a good enough policy into enough unseen homes.
So when a team still on cloud inference tells you their architecture is fine, the honest translation is often: we haven't shipped far enough to hit the wall yet. The wall is real. Their odometer just hasn't reached it.
Latency you can sometimes design around. A fixed delay is a known quantity. But jitter — delay that jumps around unpredictably — is the killer, because the robot never knows when the next instruction is arriving. A house with bad internet is, functionally, a jitter machine. And a robot reaching for a cup needs to know where the cup is now, not where it was half a second ago.
Naturally, they looked for an escape hatch. If the home network is the problem, route around the home network. They tried Starlink.
They quickly learned that in dense communities with trees, finding the 100-degree field of view to the sky is surprisingly difficult. The satellite dish needs a clear window to the heavens, and suburban tree canopy plus close-packed rooftops don't provide one. The bypass failed for the same reason the primary path failed: the physical world of the actual customer refuses to match the assumption.
The normal Wi-Fi story doesn't get simpler when you add satellites. It gets more complicated.
At this point the release timeline was at risk. Edge inference had been on the roadmap — but only as a plan-B for the Beta, a someday item. The forcing function moved it to plan-A overnight.
One engineer reportedly locked in for two days and shipped blazing-fast edge inference code straight to production. It reportedly saved the release.
And notice what that fix actually reveals, because it's easy to draw the wrong lesson. Sunday apparently already had enough local compute on the robot. What they didn't have was software that used it. The two-day fix wasn't new silicon — it was making the model run efficiently on hardware the robot was already carrying.
That distinction matters. Edge inference is a software-and-architecture decision, not merely a bigger-chip purchase order. Owning on-device inference is not the same thing as owning the on-device compute vendor — that vendor-capture question is a different fight, and I'll leave it to the people arguing about silicon supply. The point here is narrower and more durable: the moment your robot is good, the inference has to be on the machine.
After this, the team's conclusion hardened: edge inference is not optional for home robots.
The verdict is spreading past one company. One prominent roboticist reportedly put it bluntly — it will be hard to be a cloud-based token merchant, in the OpenAI and Anthropic mould, with much of robotics. The API-for-a-fee business model that minted the language-model giants may simply not port to machines that have to act inside physical space in real time.
That doesn't mean the cloud goes away. The architecture teams converge on is a split: the robot carries the work that needs an immediate answer — perception, planning, safety, movement — and the cloud keeps the work that tolerates delay, like training, fleet learning overnight, and pushing updates. The closer a decision sits to motion, safety, or private sensor data, the closer it must sit to the robot.
Now zoom out, because this is not a Wi-Fi story. It's a variance story.
Look at the RSS quadrotor result reportedly out of Zurich. Learning-based controllers usually train in large-scale simulation and then freeze at deployment, forcing conservative safety margins to survive real-world uncertainty — wind, hardware wear, battery depletion. The policy is fixed. The world is not.
Their alternative learns in the real world directly. No precise system identification, no offline sim-to-real transfer. The drone reportedly triples its speed from 2 to 7 m/s within 100 seconds of flight, pushing itself to the saturation limit of its own control interface. It reportedly adapts within a single iteration to a 30% added payload and clipped propellers, and cuts an inspection mission's time by 40% after two minutes of real-world training.
The through-line to the home robot is exact. A lab-fixed policy and clean assumptions are not a neutral starting point — they are a deployment liability. The physical world's variance is the actual adversary. Not the mean case. The variance.
Cloud inference is a bet that the network mean is good enough. Real homes are defined by their variance — the dead bedroom, the choked uplink, the neighbor's router, the tree between the dish and the sky. A policy that assumes the clean case is a policy that fails exactly when a stranger is watching it fail in their own living room.
The correct response, in the air and on the floor, is the same. Move adaptation and inference onto the machine itself — the one piece of the system that is actually present when reality diverges from the plan.
The robot got good. That's precisely why the Wi-Fi became the problem. And the only fix that survives the customer's house is the one that doesn't depend on the customer's house at all.