Build updates from Cereal.
Progress from prototype to working robots: demos, milestones, and the thinking behind them, straight from the team.
At home, the camera is locked until you open it.
A robot with a camera in your home should not be a window for anyone else, including us. So the lock lives on the robot itself, where no server and no one on our team can lift it.
Stop wins, from anywhere.
A robot in a home has to stop when asked, every time, from whatever is in your hand. This week every way to stop our robot got faster, and the robot got stricter about stopping itself.
Cereal OS left the simulator. It takes orders from one person.
Since October, Cereal OS has run our four-legged robot in a real home. Its owner talks to Cole by phone or WhatsApp, and the robot does what its body and its checks allow. Here is what it does today, and what it does not do yet.
The driver went open source. The person who knows what to print did not.
A twenty-author team released a navigation brain this month: open weights, trained entirely in simulation, the same unchanged model driving two different robot bodies. That is the layer under the site mind commoditizing on schedule. Here is what it changes, and what it cannot.
Robots get hired where someone has to sign the log.
Before a robot is allowed to work in a regulated building it has to answer four questions: who is it, where may it go, who does it call, and what did it do. Body makers do not sell those answers. Model companies do not either. That layer is the site mind, and it is why the first floors are the ones where accountability is already mandated.
Perfect motors were handed to the best minds. They scored 16.8%.
A multi-institution benchmark published this month gave frontier models flawless motor control in simulation and asked them to run long-horizon household tasks. The best model completed 16.8 percent. The motors were never the missing layer.
Amazon open-sourced the climb. The hands were human.
In one fortnight, Google started renting a whole-body brain to the best bodies in the West, and Amazon's frontier lab open-sourced a humanoid that climbs ladders, demonstrated on household chores. The motor layer is becoming a merchant commodity in public. Neither release says a word about memory, permissions, or who owns the home.
A frontier model meets a clapperboard
Someone published the complete traces of frontier models driving a real two-armed rig at a one-verb task: lift a clapper stick, let it snap shut. Every frame, every decision, every failure, public. The scores measure the distance between reasoning about a scene and acting dependably in one, and that distance is the layer.
The loss went down. The robot still failed.
A careful failure study published this week shows a world model's training loss can improve while the planner on top of it stays at chance. It is the strongest outside evidence yet for the bar we hold ourselves to: judge a robot by certified outcomes and attested actions, never by its training curves.
Persistence became a benchmark. Here is ours.
The research field has started scoring what a robot believes, not just what it completes. We think that bar is still too low for a machine in a family's home, so here is the three-clause test we hold ourselves to, published as a request for comment.
The Android shipped. The account didn't.
Everyone is building the operating system for robot control, several times over. Nobody is building the account that logs in: whose machine this is, what it remembers, what it may do. The OS for control is crowded. The account for the home is still open.
Agents can write robot skills now. Someone still has to decide if they're real.
A new wave of systems lets LLM agents write robot control code from a prompt. Directionally right, and missing the half that matters in a home: certification. As of today, cereal-os v0.2 lets any frontier agent author skills, and the twin decides if they are real.
The bottleneck was never the body. It isn't the brain, either.
WAIC settled a new consensus: the hardware is ready, the intelligence is not. Both halves still miss the part that makes a robot yours. Capability makes a robot useful. Continuity makes it yours.
What phones teach us about robots
Phones split into hardware and OS, and the OS carries your life between bodies. Robotics has not made that split yet. Cereal is building the layer that does.
Everyday Robots was working. It just never shipped.
Google's Everyday Robots was doing real work inside Alphabet when it was shut down in January 2023, before it ever shipped to the world. Our founders helped build it. Cereal is the regroup to finish the job.
Simulation in the loop: the Everyday Robots patent behind our world loop
US11938638B2 is usually filed under sim-to-real. It is really about keeping a live simulation inside every control cycle of a deployed robot. Here is what it actually says, and how Cereal builds above it.
The mind was never the machine
The intelligence that runs a place never lived in any one machine. Bodies, arms, and rovers come and go; the mind that knows the place stays. That is the whole thesis.
Cereal OS v0.1: one mind, two bodies
The first running version of the portable mind: the same intelligence and memory completing the same task on two different robot bodies, in simulation, with every action attested.
The brain already exists: meet Cole
Most teams are still building the mind. Cereal is dropping in one that already works, and giving it a body.
Why we build the mind for life, not labor
Most humanoids are being built for the factory floor. Cereal starts with the home, the people, and the moments where presence matters most.
Introducing Cereal: one mind, any body
AI is leaving the screen. We are giving the assistant that already lives in your life a body.