Kittle Industries
Training the machines that will build on the Moon.
The thesis
Frontier AI is the mind. We train the hands.
Robot brains are now built in two layers: a slow planning layer that decides what to do, and a fast skill layer that turns intent into motor commands many times a second. Physical AI is short of data for both. We produce that data, and the environments to train on, for places people can't go: lunar and Martian construction sites.
The skill layer
Moment-to-moment control: how hard to crowd a bucket into regolith, when to lift, how to keep a machine from stalling or tipping. It's learned through practice, the way a veteran operator learns, and practice is what a game supplies.
The planning layer
Where to dig, where to dump, which route to drive. On Mars, signal delay forces supervised autonomy: a person lays out the plan and the machine carries it out. Plans paired with their outcomes are training data for the planner.
Superintelligence can plan a moon base. It can't dig one.
Why off-world construction
The niche the big players skip, where the space money is going.
Data is the bottleneck
Human demonstrations with real machines and operators are slow and expensive to collect. A game turns that cost into play.
Nobody collects it
General robotics companies build for homes, warehouses and roads. Lunar excavation is too specialized for them to collect themselves, so they're likelier customers and partners than competitors.
Latency shapes the data
The Moon is about 1.3 s away each way, so delayed teleoperation works. Mars is 3 to 22 minutes away, so it needs supervised autonomy. That gives two kinds of data: control under delay, and plans for autonomy.
Products
One platform, two customers.
Offworld Builder
A game that trains the robots that will build on the Moon. Players run an excavator on NASA terrain of the Moon and Mars, in the browser. Every session is recorded as robot training data.
Askesis
Physical skill, learned through practice. Demonstration datasets and environments for teams building autonomous construction machines, in formats robotics teams already use.
The loop
Play, learn, refine, plan, deploy.
No one else runs this whole loop for off-world construction. Here is where each step stands today.
- Live
Play
Different players solve the same job differently. That variety is the point.
- Working
Imitation learning
Behavior cloning with DAgger on logged sessions. Working today, trained on scripted-expert data only.
- Planned
Reinforcement learning
Headless batch simulation with rewards for progress, energy and safety, to refine policies past their teachers.
- Partly built
Mission plans to autonomy
People lay out dig and dump zones and waypoints; the machine executes. The plan editor is in the game; training a planner on it is next.
- Planned
Sim-to-real
A small physical robot with an exact simulated twin, running the same trained policy in both.
Research
Physics checked against NASA measurements.
Reinforcement learning exploits errors in the simulator, so fidelity is a requirement, not polish. We test the models against published experiments and report where they miss.
Excavation force
Median of predicted to measured force on 5 NASA Glenn bucket tests in lunar simulant (range 0.76–1.19×).
Wheel drawbar pull
Against a NASA Glenn single-wheel test at 30% slip and above. About 30 N low below 30% slip.
Learned policy to spec
Held-out starts where a policy trained on game sessions digs the trench to spec. Scripted-expert data only, no human data yet.
Get in touch
Building for the Moon or Mars?
We're talking to robotics and space teams about data and custom environments, and to investors and grant programs about what comes next.