Stage 01 Shipping
The brain module
A plug-in module that makes a machine a person drives into one that drives itself. The installed base becomes the fleet, so volume does not wait on new hardware.
Two stages. The brain, then the machine.
We start with outdoor working machines: the work with the clearest unmet demand, and the hardest conditions to do it in. Both stages have to survive manufacturing tolerance, cost-down, service and warranty — not a demonstration on a good day.
Stage 02 In design
Our own robot
A purpose-built working machine with a tool ecosystem, specified and verified in simulation before anything is fabricated.
It fits the machine you already own.
Sensing, compute and an independent safety chain in a single module.
- Reads the site on arrival, sets its own boundary, plans its route, finishes the job. Empty seat.
- No survey, no buried wire, no route drawn in advance.
- Nothing about the host machine is modified.
- One architecture from ride-on mowers to golf carts, utility vehicles and tractors — only the actuation layer differs.
A whole course, running as one system.
Developed with our manufacturing partner.
- Autonomous mowers, carts, delivery, washing and charging robots across an entire golf course.
- The most complete outdoor deployment the stack has had, and where much of the platform came from.
- Our stack also runs inside a vehicle line that is first worldwide by volume in its segment — roughly half the global market, and around 800,000 machines delivered.
Machines of our own.
Designed inside the loop, before anything is fabricated.
- A purpose-built working robot with a tool ecosystem.
- The first machine whose design is settled by simulation rather than discovered on a prototype.
- Then the same brain in warehouse, service and domestic machines. The world model changes; the loop does not.
Robots are where chips were before EDA.
A robot's bill of materials falls every year. Its development cost does not. That is what keeps robots out of volume, and it is the problem we actually work on.
Every robot is a prototype
Perception, control and safety get rebuilt for each machine. Almost nothing carries over.
Testing needs the real world
Failures are found by driving a physical machine into a physical place. Slow, unrepeatable, gated by weather and season.
Volume is where it breaks
Manufacturing tolerance, cost-down and warranty all change the machine. Every change reopens the testing.
Chips solved this. Cars solved this.
- Silicon got cheap when design moved into simulation: verify first, fabricate once.
- Automotive followed, with closed-loop validation before the hardware exists.
- Robotics has no equivalent.
That is what we build.
Any robot that has to work in a place.
The loop is domain-agnostic. Only the world model changes.
Outdoor working machines
First. The hardest case.
Warehouse & logistics
Structured, dense, high duty cycle.
Service & inspection
Mixed indoor and outdoor, people present.
Domestic & indoor
Cluttered, unmapped, endlessly varied.
One loop. Capture to machine, and back.
Development moves into simulation. Each stage produces exactly what the next consumes, and machines in the field feed the first stage again.
- 01 Capture A phone or drone pass over a real site.
- 02 Cousins One layout becomes unlimited worlds.
- 03 Train & verify Closed-loop learning against work that changes the scene.
- 04 Configure The whole onboard system, not just the policy.
- 05 Deploy Same specification drives the real machine.
Field captures return as new worlds
We built it for ground that will not hold still.
The ground moves
Soil deforms, grass compresses, slopes change what the machine can do.
The light never repeats
Hour, season and weather rewrite every image the robot sees.
The work rewrites the scene
The machine changes the place it is working, then has to perceive the change.
No map survives
Sites grow, get cluttered and get rearranged between visits.
One clip becomes unlimited worlds.
A phone walked across a site, or a drone flown over one. That is the entire input.
- A twin reproduces one place. A policy trained on it learns that place, and the next site defeats it.
- A cousin keeps the structure and re-rolls everything else: species, materials, light, weather, growth.
- The family transfers. No single member does.
N cousins
from one capture
Scene count beats scene fidelity.
The world remembers what the machine did.
Working robots change the scene they are looking at. They have to see what they have already done — and no navigation simulator represents that at all.
- A world-aligned record of work state runs through the simulation.
- The machine edits it. The cameras see it. The reward reads it.
- One mechanism, different channels — so the second task is not a second system.
- Cut height & direction
- Mowing, striping, verge work
- Trimmed shell
- Hedging, edging, pruning
- Debris density
- Clearance, sweeping
- Cleared area & object state
- Collection, pick-and-place, tidying
03
Train & verify
- Dynamics and photorealistic rendering advance in lockstep.
- Policies learn inside a predictive model, evaluating outcomes several steps ahead.
- Sensors modelled by failure mode, not datasheet.
- Deterministic. A failure becomes a regression test.
04
Configure the whole system
- Perception, planning, middleware, safety chain and compute budget as one object.
- The trained network is a fraction of what ships.
- Change one layer and the loop prices it — success rate, latency, watts.
05
Deploy
- Simulated and physical machine from one specification, so they cannot drift.
- Policies verified in the loop have gone onto real robots with no tuning.
- Layouts that appeared nowhere in training.
Then back again.
Machines in the field return captures of the places they work. Those become new cousins, which train the next policy, which goes back onto the machines. The loop has no end. The loop is the product.
Every layer is a seam. Nothing is welded shut.
Four contracts define the system. Each is a place where a different model can be substituted without disturbing anything above it.
SceneSpec
Terrain, surface materials, vegetation, objects, sky — and the region to be worked.
RobotSpec
Kinematics, mass, actuation limits, sensor mounts. One description, two consumers.
ObsSpec
Render passes, sensor channels, noise models. Fixed layout, so a policy cannot tell which engine produced a batch.
WorkStateSpec
What the work did to the world.
Each layer runs today on a model good enough to train against. Each would rather run on one that is right.
Sensor physics
Today
Render passes with fitted degradation
What can replace it
Physics-based cameras, lidar, radar and thermal — true optical and multispectral behaviour
Environment physics
Today
Contact and terrain models tuned to measured behaviour
What can replace it
Soil and granular mechanics, deformation, thermal, weather, electromagnetics
Machine physics
Today
Rigid multibody dynamics, measured actuation limits
What can replace it
Structural response, drivetrain, motor and battery across the real operating envelope
Embedded software
Today
Hand-built stack, verified as a system
What can replace it
Safety-certified code generation and on-target verification
Run the loop backwards.
- Today the loop answers one question: does this system work in this world?
- It already holds most of what it needs to answer a better one: what machine should exist?
- Which sensors earn their place. How much torque on that slope, in that soil — and which motor delivers it. Which compute module the stack fits inside, and what that costs in battery.
We have not built this. The architecture was built so that it can be — and the layers it would need are exactly the ones left open.
Cambridge, then cars. Now robots.
-
Where we started
Founded at the University of Cambridge by scientists and serial founders, out of research on how machines locate themselves and move through the real world.
-
What we did next
Selected by one of the world's three largest electronic design automation vendors to build vehicle-level closed-loop simulation inside its pre-silicon verification platform — where driving software and the silicon beneath it are validated together, before either exists in hardware.
-
Why we are here
Robotics has the same bottleneck cars had, and none of the tooling. We are moving the experience across, and building the machines with it rather than only selling the method.
Per scenario
Hours → seconds
System-level verification of one driving scenario, end to end.
Scenario coverage
Unlimited
Generated, not driven.
End customers
Top three
Three of the world's largest automakers and their tier-one suppliers.
- Founded at
- University of Cambridge
- Team
- Scientists and serial founders · Cambridge alumni
- Experience
- 10+ years each in autonomy, robotics and simulation
- Backing
- Earlier ventures backed by Cambridge University's venture arm and leading deep-tech investors
- Research
- Published across the major robotics conferences and journals
Get in touch.
Or write directly — contact@xcrobotics.ai