Foundation
Physical AI needs a different foundation.
Tokens don't have mass. Weights don't account for watts. The stack your robot runs on should start from the physics, not end there.
Math-grounded, not token-grounded
Every primitive has a closed-form math model. Safety envelopes are derived, not trained. Controllers trace back to their Lyapunov function. No emergent behavior where you needed a proof.
Picojoule-metered
Energy receipts at every call site. auto-grade 7 nm cost model baked in. Know what your pipeline costs before you ship it to a battery-powered drone.
80 primitives that compose
mgai-fuzz feeds mgai-safety feeds mgai-auto feeds mgai-fleet. Same workspace, same types, built together. The composition is the product.
One install. Native + WASM.
One cargo add and the whole substrate is on your workbench. The same code runs on Apple Silicon, a Jetson Thor, and inside a 398 KB WebAssembly module.
The substrate
Eighty primitives. One workspace.
Each primitive is a real Rust crate in a single workspace. They share types, test infrastructure, and energy metering — so a safety envelope from mgai-auto plugs directly into a fleet coordinator from mgai-fleet without glue code.
Metering
Every call, a picojoule receipt.
Representative numbers from the integrated_physical_ai reference run — 400 pipeline steps, mobile manipulator under RSS safety + DAM-VLA routing + digital-twin divergence, composed in one trajectory.
Go
See it. Run it. Build with it.
# add the kernel + a vertical of your choice
cargo add mgai-core mgai-auto mgai-vla mgai-digital-twinRust 1.94 · edition 2024 · native + wasm32-unknown-unknown