MaximumRobotics
The frontier was trained on a narrow sample.
We measure the rest of the world.
H(X) — m-e-r.xyz — telemetry begins Q4
The thesis
Models are what they eat.
Every capable machine intelligence of this decade was trained on the same narrow slice of human experience: clean homes, ordered warehouses, staged kitchens. The sample is biased, and every capability ceiling traces back to it.
The world is high-entropy.
Real work happens in crowded markets, monsoon light, improvised tools, a thousand dialects of the same task. This data exists nowhere in any corpus — because nobody has been standing where it happens.
Sampling is the company.
We are building the instrument and the network to sample human work where it actually occurs — at a fidelity and scale the frontier cannot generate, cannot buy, and has not seen.
THE DERIVATION
The derivation is 150 years old.
Maximise entropy subject to constraints and you do not get chaos. You get the Boltzmann distribution:
pi = e−βEi / Z
The law that falls out of a gas, a crystal, any system in contact with reality. Jaynes showed in 1957 that it is not a trick of physics — it is the least-biased description of any system you do not fully know. Our constraint set is the world itself: its markets, its light, its improvisation.
Solve for pi.
OBJECTIVEmax H(p)
CONSTRAINTΣ pi = 1 · the real world
RESULTpi = e(−βEi)/Z
argmax H s.t. reality
pi = e−βEi / Z
β = the world's constraints
Z = everything we haven't seen
The instrument
You know what maximum entropy does in a loss function.
We build the data that does it to your policies.
We built a device for measuring human work. Compact. Head-worn. Silent. It records the one thing no simulation produces — the true statistics of hands, objects and intent, in environments that were never designed to be training data. Diverse scenes. Diverse hands. Diverse ways of solving the same task — which is precisely the variation your entropy bonus is starving for.
That's all we say about that for now.
Full specification shared with qualified partners.
The network
An instrument is useless without a network to deploy it into.
We are onboarding a distributed measurement workforce across high-entropy geographies — places where the distance between "a task being done" and "a task being recorded" is currently infinite.
Every sampling site is chosen for one property: maximum divergence from every dataset the frontier already owns.
sites live: ▓▓░░░░░░ · cohorts forming
Questions
What is Maximum Entropy Robotics?
A company that measures real human work — head-worn egocentric capture, deployed through operator networks — to produce manipulation training data with diversity the frontier cannot generate.
Why "maximum entropy"?
In reinforcement learning, an entropy bonus keeps policies exploring instead of repeating. In physics, maximising entropy under constraints yields the Boltzmann distribution — the least-biased description of any system. We constrain the maximisation with the real world, and sample what falls out.
Who is it for?
Teams training vision-language-action models and robot policies who need out-of-distribution egocentric manipulation data — consented, stereo, hardware-synced, and delivered non-exclusive in RLDS/LeRobot formats.
When does sampling begin?
Instrument deployments begin Q4 2026. Qualified partners can request the technical specification now.
The invitation
Three kinds of people will read this page.
If you train models — your entropy bonus can't help a policy that's never seen the tail. We'll show you what the tail looks like.
→ [ request the technical brief ]If you work with your hands — in the places we'll go first — you are not a data point. You're the measured world, and you'll be paid like it.
→ [ join the first cohort ]If you back hard problems — the scarcity we're addressing is measured in orders of magnitude, not percentages.
→ [ read the thesis ]