The chip on the robot's side
A facility's inference pool can serve perception, decisions, and adaptation. Control, motion, stops, and safety stay on the machine. BramHA is being designed for the machine's side of that line.
· 4 min read
IN SHORT
A robot's work splits in two. Perception, decisions, and adaptation are heavy and can be served from inside the building by a Czar Grid. Control, motion, stops, and safety cannot wait and cannot fail, so they stay on the machine. BramHA is being designed for the machine's side of that line. It is in development and no chip exists.
A robot on a working floor does two kinds of work at once. Some of it is thinking: making sense of what its cameras see, deciding what to do next, adapting to a new part or a changed layout. Some of it is acting: holding a joint where it should be, moving it on time, stopping when it must.
BramHA is Element 31's silicon for the second kind. It is in development and no chip exists. This article explains the line between the two kinds of work and why the program sits on the machine's side of it.
Two kinds of work
The thinking work is heavy. Perception, decisions, and adaptation run on large models, and large models need dense, power-hungry compute. That compute is costly to put inside every machine, and it ages quickly as models change.
The acting work is light by comparison, but strict. A control loop has to close on time, every time. A stop has to work whether or not anything else does. This work is judged by whether it happens when it should. How much computing it takes is a secondary matter.
The two kinds of work want different hardware, in different places.
What the pool serves
Element 31's answer for the heavy work is a Czar Grid: several Czars working as an inference pool that serves every machine in a facility, from any vendor, over the site's own network. The pool serves perception, decisions, and adaptation from inside the building.
Sharing that compute has a practical logic. A machine does not need a large model at every moment, so many machines can draw on one pool. And the pool can be changed without opening a robot.
The Grid is not a controller. It does not sit inside any real-time or safety loop. It answers what a machine asks; it does not move the machine.
What stays on the machine
Control, motion, stops, and safety stay on the machine as its vendor built them. This is the work that cannot wait and cannot fail.
The reason is simple. A network link, however good, can be slow for a moment or drop altogether. A machine whose joint control depended on that link would be only as dependable as the link. So the loop closes on the machine, the drive sits at the joint, and the sensors are read where they are measured.
If the pool cannot be reached, the machine still controls itself and still stops. What it does while it waits for an answer is for its vendor to decide. A control loop that closes on the machine covers why this work stays local.
Why the split decides the chip
Once the heavy AI work is in the pool, the machine's own silicon has a narrower list of jobs: control, motion, sensing, and identity. Those jobs need certainty, not density. None of them needs the leading edge of chip manufacturing, and mature process nodes are enough for them.
That is why the first design is deliberately small: a real-time controller with a root of trust. It is not a vision processor and it is not an AI accelerator.
The trade should be stated honestly. A machine built this way relies on the building's pool for its heavy AI work. A machine that has to carry all of its intelligence on board, with no pool to call on, needs denser silicon than this program is designing. That is a different problem, and BramHA is not aimed at it.
What the split does not change
BramHA does not move the line. The Grid still never sits in the loop. The silicon is a part a robot maker may build into the machine, and the machine's safety is still the vendor's design and the vendor's responsibility.
The silicon is not a safety system, and no safety or functional-safety certification is claimed for it. The Grid thinks, the machine acts sets out how the two sides relate.
Nothing described here has been built. No chip has been taped out, the specification of the first design is not published, and nothing can be ordered. The page at /program lists what is done and what is not.
Questions this raises
- Why not put all the computing inside the robot?
- Large models need dense, power-hungry compute that is costly to place in every machine and that ages quickly. A pool in the building can be shared across many machines and changed without opening a robot. The work that cannot wait still stays on the machine.
- Does the inference pool ever control the machine?
- No. Czar Grid serves perception, decisions, and adaptation over the site's network. It is not a controller and does not sit inside any real-time or safety loop.
- Does BramHA run AI models on the robot?
- It is not being designed to. The first design is a real-time controller with a root of trust, not a vision processor or an AI accelerator. Heavy AI work stays with the pool in the building.
KEEP READING