OpenIE · Physical AI Library · Rust
Math-Ground AI.
Physical AI that tells you what it just cost.
# coming to crates.io — not published yet
cargo add mgai-algebra mgai-core mgai-vlaThe wound
Language AI rose on the internet of words.
But the world isn’t made of words. It’s made of forces, motion, sensors, and joules.
The job
Physical AI answers to the world — every frame.
A robot that doesn’t respect physics crashes. A drone that doesn’t respect energy dies. A car that doesn’t respect safety kills.
The discipline
Closed-form, wherever it can be.
Safety envelopes derived in closed form where we have them — RSS following distance, control-barrier checks. Controllers traced to a Lyapunov function are the goal, not yet the code. No emergent behavior where you needed a proof.
The receipt
Every receipt says how it knows.
You know what your pipeline costs before you ship it to a battery-powered drone. Not after.
A simulated 400-step integrated run on a mobile manipulator — RSS safety, DAM-VLA routing, digital-twin divergence — books 87.9 joules against fixed per-stage energy budgets. That is a model, not a meter reading; the receipt below is step 207. On real silicon, mgai-meter reads the CPU package counter instead: 2.40 J ± 0.026 J on a Ryzen AI Max+ 395 and 4.88 J ± 0.052 J on an Intel + RTX 4050 laptop — measured, whole-package energy for its ~0.1 s run.
- Plan 2.0 ms · 200.0 mJ
- Actuate 80 µs · 10.0 mJ
- Perceive 200 µs · 5.0 mJ
- World model 150 µs · 3.0 mJ
- Sense · control · safety · telemetry 95 µs · 1.8 mJ
- Per-step total 219.8 mJ
Evidence: modeled — fixed per-stage budgets from a simulated run, not a meter reading.
The shape
Five atoms. Eight compositions. One workspace.
Safety feeds auto feeds fleet. Perception feeds world-model feeds plan. 112 crates in one Rust workspace, built on one small algebra. Metered where we can measure; labelled where we can’t.
The reach
It runs in your browser.
WebGPU. About 520 kB to download, 180 kB compressed. Joints move. Joules tick. No cloud, no wait.
The lineage
The pattern flew on NASA’s Deep Space 1 in 1999.
Plan, execute, diagnose. Verify by conflict, not per-claim guess. Anytime semantics. Decisions reported with provenance. We picked up where they left off.
math ground dot AI
Built on math.
For the physical world.
# coming to crates.io — not published yet
cargo add mgai-algebra mgai-core mgai-vla