Physical AI Research

Intelligence that acts in the physical world.

Physical AI is the science of machines that sense, decide, and act under the constraints of reality: gravity, contact, noise, and time. The Institute's research spans five areas across perception, control, and coordination.

Research topics

27 open lines of work.

Each is a standing question with the evidence attached: a technical report where one exists, usually a runnable simulation, and a plain statement of what is still unsettled. Read them as teaching material, because that is what they are written as.

TR-2026-26Building the energy compute futureSwap-2C · the barTR-2026-25The energy-first turnSwap-2C · the mapTR-2026-06MathGround: joules, not tokensSwap-2C · energyTR-2026-07Provable by constructionSwap-2C · proofTR-2026-08OmniSense: perception as a volumePerceptionTR-2026-09The contact layerPerception · touchTR-2026-22Touching soundPerception · soundTR-2026-02Ternary Physical AIComputeTR-2026-03Thermodynamic Physical AIComputeTR-2026-01The open field of computingComputeTR-2026-28Quantum information at the edgeCompute · quantumTR-2026-10Spatial AIPerception · world modelTR-2026-11The Computable World ModelWorld modelTR-2026-12Graph of the WorldWorld modelTR-2026-13Spatial RFPerception · RFTR-2026-14The surface that pays twiceMaterialsTR-2026-15The printed bodyEmbodimentTR-2026-16VLI: built to interactEmbodied AITR-2026-27Space logistics & transportationAutonomy · spaceTR-2026-29Physical AI in spaceAutonomy · space · economicsTR-2026-30Physical AI in the fieldAutonomy · field operations · economicsTR-2026-20The matching principleEmbodied · contactTR-2026-21The adaptive unitLearningTR-2026-18The Human LayerWorkforce · competencytopicPerception, Privacy, and PolicyWorkforce · sensingTR-2026-19The Diffusion LayerWorkforce · siliconTR-2026-31The future of commerceWorkforce · commerce

The papers behind them →The lab that runs themThe Control Benchmark standingsFerralloy · the edge-deploy layer

Open release

Weights you can run.

Our releases ship the weights, the dataset they learned from, and a way to actually run them — on your own device, no server.

World modelLearned simulator · playable

nano-world-model

A simulator with no physics engine: a small neural network learned momentum, walls, and the bounce from watching, and now paints every next frame from your action — feeding its own output back. Drive it with the arrow keys.

Play the dream →
VLAOpen-vocab · perspective camera

nano-vla-arm-3d

An open-vocabulary vision-language-action policy running its whole forward pass on-device. Type any instruction — a real language model embeds it on your device — and it drives an articulated arm's joints, from a perspective camera view, to the block you named. 119,986 parameters.

Drive it with words →
CheckpointOn-device · CC-BY-4.0

forge-arm-reach-bc

Our first released policy checkpoint — 17,923 parameters, 72 KB. The exact published weights close the loop live on the same MuJoCo arm they were trained on, at ~100K inferences per second. No cloud.

Run it on the page →
Grown circuitsMorphogenesis · checked by physics

describe a rule, not a design

A base case and a recursive rule grow an adder of any width, with no bus size ever declared. The grown graph is lowered to transistors and solved as one nonlinear circuit, so the analog voltages decide whether it computes. Verified from 1 to 16 bits, 480 transistors.

See what grew →
CertificatesWorst case · named region

certificates and refusal

A safety benchmark counts how often a robot declined an unsafe action. A certificate names the region where refusal is guaranteed, the recover that holds it, and what holding it costs. Three barriers certified as worst cases on an open arm model: the work surface, a person, and the arm's own body.

Read the method →
EnergyExplorable · runs on your device

the energy lab

Computing's next limit is electricity, and the reason is not the arithmetic. Watch a calculation pay for itself one operand at a time, then watch a hundred coupled variables fall into an answer without searching for it. Two simulations, the physics behind the substrates being built to escape the memory wall, and a map of eighty-five organisations attempting it.

Run the simulations →
BenchmarkOpen · conservation-law fidelity

physics-fidelity benchmark

The axis appearance hides: does a model obey physics? Eight probes — energy, momentum, contact, friction, identifiability, conservativeness, dissipation and structural prediction — score any dynamics model against analytic ground truth, verified on the real SO-101. A trained black-box world model fails; a structured one, same data, passes. Structure beats data, measured.

See the benchmark →
PhysicsSoft bodies · grains · on-device RL

soft bodies, grains & on-device control

The rest of the pure-Rust physical-AI stack, in the page: a volumetric Neo-Hookean jelly wobbling under gravity, discrete-element grains pouring into a settling pile, and a cartpole that swings itself upright by rolling out hundreds of futures in parallel. Every solver verified against a physical invariant, all on the device in front of you.

Run it on the page →
SensorsVerified · browser-native · on-device

the verified browser sensor renderer

Synthetic depth, segmentation, and lidar — the sim-to-real currency — rendered by a pure-Rust signed-distance ray marcher compiled to WebAssembly. Sensor simulation that renders and proves itself, in the browser, on your own hardware, free to anyone — no cloud GPU. The tracked sphere's depth is checked against the exact ray–sphere solution every frame.

Run it on the page →
FluidsVerified · differentiable · in-page

the verified browser fluid bench

A pure-Rust CFD stack — MAC projection and lattice-Boltzmann — compiled to WebAssembly and run in your browser. A lid-driven cavity grades itself against the Ghia (1982) reference table as it converges; a gust bench shows the coefficient model break and the resolved solver reveal the fix; a self-propelled swimmer earns thrust from nothing. Every panel renders the flow and proves itself.

Run the solver →
Skill layerOpen · cross-vendor · safety-enveloped

skillpack

The open skill layer: a robot skill is source you own — a policy, a typed contract, a capability manifest, a safety envelope. It gates a skill against your robot's declared capabilities before anything moves, and clamps every action to a runtime envelope. Pick a robot, watch the registry gate skills against it, then corrupt the policy and watch the envelope hold — live in your browser.

Open the live demo →
Join us

Become an investigator, or intern.

The Physical AI Investigator Program turns an early-career researcher into an independent PI. Students can intern with a lab, and investigators who hold their own funding can bring their program here.

Physical AI Investigator Program →

Or see all open roles →