The frontier was trained on a narrow sample.
We measure the rest of the world.

Calibrated stereo capture of real human work. 3D hands, measured, built to your spec.

H(X) — m-e-r.xyz — cohorts forming

p(x)uniform
01

The thesis

Models are what they eat.

Most robot policies are trained on a narrow slice of human experience: clean homes, ordered warehouses, staged kitchens. Collapsed samples produce collapsed policies.

The world is high-entropy.

Real work happens in crowded markets, monsoon light, improvised tools, a thousand dialects of the same task. Most of it has never been recorded, and what has been recorded was rarely measured.

Footage is cheap. Measurement isn't.

Millions of hours of egocentric video now exist, much of it free. Very little of it measures the hands in 3D, with known geometry, and that is the part that transfers to a robot.

Sampling is the company.

We build the instrument and the network to measure human work where it actually happens, to your specification.

02

The derivation is 150 years old.

Maximise entropy subject to constraints and you do not get chaos. You get the Boltzmann distribution:

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 a system you do not fully know. Our constraint set is your specification and 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

p(x)uniform
03

The instrument

You know what maximum entropy does in a loss function.
We build the data that does it to your policies.

We're building a device for measuring human work. Compact. Head-worn. Silent. It measures hands and objects in 3D, 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.

Cameras
4 × 720p @ 30 fps
Geometry
Forward stereo pair (scene) · downward stereo pair (hands)
Sync
Hardware, < 1 ms, per-frame timestamps
Calibration
Per-unit intrinsics and extrinsics, shipped with every dataset
Delivery
LeRobot, RLDS or your loader

Every delivery ships with its 3D hand error in millimetres, measured against a multi-camera reference.

Why stereo: in Meta's HOT3D benchmark (CVPR 2025), going from one headset camera to two cut 3D hand-keypoint error from 15.4 mm to 10.9 mm.

The rest is in the brief.

Full specification on request.

p(x)uniform
04

The network

An instrument is useless without a network to deploy it into.

We recruit and train operator cohorts where the work is, starting in Central Java, Jakarta and Colombia.

Every site is chosen for divergence from the public corpora, and we add sites wherever your specification needs them, including where your robots will ship.

Cohorts, not crowds. A programme runs on a trained cohort of about 25 operators, not thousands of strangers.

sites: Central Java · Jakarta · Colombia — cohorts forming

05

The offer

Start with a pilot.

Send us your task list. We capture 25 hours of it in our environments, calibrated to your robot's camera, and deliver it in your format with a 3D hand-accuracy report. You run it on your own evaluation.

Then a programme.

Accepted against written acceptance gates: hands in frame, sync, 3D error ceiling, composition, consent. Exclusivity windows available.

[ request a pilot ]
p(x)uniform
06

Questions

What is Maximum Entropy Robotics?

A company that measures real human work. A head-worn stereo instrument, deployed through trained operator cohorts, produces calibrated 3D hand–object data from environments public datasets don't cover, built to each lab's specification.

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 a system you do not fully know. Your specification is the constraint. The world supplies the entropy.

Who is it for?

Teams training VLA, world-action and robot policies who need human manipulation data they can't get from public corpora: consented, stereo, hardware-synced, calibrated to their robot, and delivered in LeRobot, RLDS or their own loader. Bespoke by default; exclusivity windows available.

How is this different from free egocentric video?

Open corpora are large but mostly monocular and low-resolution, with hands estimated after the fact. We measure hands in 3D at capture, with known geometry, and report the error on every delivery.

When does sampling begin?

Cohorts are forming now. Pilot slots are open now.

p(x)uniform
07

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 a pilot ]

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 — footage is abundant; measured hands are not. That gap is the company.

→ [ read the thesis ]