Physical AI

Intelligence
that lives
in the machine.

Core On Dynamics builds an on-device AI stack for industrial robots. A fast vision-language-action policy and an accurate domain LLM run together on edge hardware, and every action passes a deterministic safety layer before it reaches the machine.

Talk to us Explore the stack
NVIDIA Jetson Thor module surrounded by robot arm, edge box, drone, quadruped and conveyor NVIDIA Jetson Thor module surrounded by robot arm, edge box, drone, quadruped and conveyor

CoreOS · Safety is architecture

Models propose.
Architecture verifies.
The system authorizes.

Safety is not left to the model. A deterministic layer written in Rust checks every action command before it reaches the machine.

Read the essay Why on-device
coreos::permit
// every action command, every cycle
match coreos.verify(&cmd, &state) {
  Ok(permit)   => actuators.execute(permit),
  Err(reason) => planner.retry_or_safe_stop(reason),
}
// checks: physical · spatial · collision · state

Illustrative only. Not production code.

CoreEdge

One module.
Many machines.

System 1 and System 2 run together on NVIDIA Jetson Thor-class edge hardware, closing the loop without a cloud round trip.

See CoreEdge Solutions
Jetson Thor module connected to robot arm, humanoid, gantry, quadruped, drone and edge box icons
The stackPerceive, reason, verify, act SolutionsShipbuilding, steel and more Lab notesThe C-Signal tech blog Pilot with usBring a real task from your line

What is Physical AI

Perceive. Reason. Act. All on the machine.

Physical AI is AI that takes actions in the real world. In a factory that loop has to close on the device, in real time, and safely. Here is how our four layers map to it.

01 · Perceive & act

CoreVLA

System 1 · acts

The fast vision-language-action policy that turns what the robot sees and is told into actions on the line.

02 · Reason

CoreLM

System 2 · adds context

A slower but accurate domain language model that adds context, planning and explanation to the loop.

03 · Verify

CoreOS

Safety · verifies

A deterministic Rust layer checks every command: physical limits, workspace, collision and system state. Fail means no execution.

04 · Act

The machine

Actuators · authorized

Only a verified command reaches the arm. Otherwise the model retries, CoreLM proposes an alternative, or the system goes to safe stop.

NVIDIA Jetson Thor module
Runs on

CoreEdge · NVIDIA Jetson Thor-class edge hardware

Both models run together on one edge computer, closing the loop without a cloud round trip.

Products

One stack, four layers, one loop

Hardware

CoreEdge

Edge compute on NVIDIA Jetson Thor. System 1 and System 2 run together on the device, so the loop closes without a cloud round trip.

Learn more
System 1

CoreVLA

Vision-language-action policy. The fast policy that turns what the robot sees and is told into actions on the line.

Learn more
System 2

CoreLM

Domain LLM. A slower but accurate domain language model that adds context and reasoning to the loop.

Learn more
Safety

CoreOS

Deterministic safety layer in Rust. Every action command passes CoreOS. The model proposes; the system authorizes.

Read the essay
Platform

Consent-based data flywheel

Anonymized fleet data improves the base models only with explicit customer consent. Task data and fine-tuning adapters are never shared.

How consent works

Solutions

Where Physical AI meets heavy industry

Robots that work on large, curved and variable parts, where the cost of a wrong action is high.

industry-shipbuilding.jpg
Focus industry

Shipyard Physical AI

Welding robots, digital shipyards and humanoids are moving from pilots into production. Our stack is built for robots that work on large, curved and variable hull parts.

  • Runs on the robot, not in the yard's network
  • Every motion verified by CoreOS
  • Customer process data stays on-site
Read our shipyard briefing (KO)
industry-steel.jpg
Focus industry

Line heating and curved-plate forming

Our public research notes look at what an AI heating plan has to respect: temperature limits, heating paths, cooling paths and the resulting microstructure of hull steel.

  • Hard limits enforced deterministically
  • Domain reasoning from CoreLM
  • Plans explained, not just executed
Steel processing notes (KO)
industry-manufacturing.jpg
Manufacturing

Robot cells that keep the line running

If the network drops or a cloud call is slow, the robot still has to work. On-device AI keeps the decision loop inside the cell.

  • No cloud round trip in the control loop
  • Predictable operating cost
Discuss a pilot
industry-logistics.jpg
Logistics

Handling variable items on the floor

A vision-language-action policy understands what it sees and what it is told, so mobile robots can handle variety without a round trip to the cloud.

  • Language instructions to actions
  • Workspace limits enforced by CoreOS
Discuss a pilot
industry-inspection.jpg
Inspection

See, explain and flag on the line

CoreVLA perceives; CoreLM adds context and explanation. Images and process data never have to leave the building.

  • On-site data by default
  • Explanations alongside decisions
Discuss a pilot
industry-robotics.jpg
Robot OEMs & integrators

Put the stack inside your machine

CoreEdge brings both models and the CoreOS safety layer onto one Jetson Thor-class computer that can sit inside arms, mobile robots and cells.

  • One module, many machine types
  • Per-customer adapters, isolated
Talk to us

Why on-device

Why factory AI runs on the device

01

Latency

Robots act on the line in real time. The decision loop cannot wait on a cloud round trip.

02

Works offline

If the network drops or a cloud call is slow, the robot still has to work.

03

Data stays on-site

Each customer's task data and fine-tuning adapters stay isolated and are never shared with another customer.

04

Safety in Rust

A deterministic layer written in Rust, not the model, checks every action command before it moves the machine.

05

Predictable cost

Running on the device keeps the cost of continuous operation predictable.

"Models can be replicated. Real-world operational data accumulated in production cannot."

Safety Is Architecture
100%of action commands pass CoreOS before reaching the machine
4deterministic checks per command: physical, spatial, collision, state

Pilot with us

Bring a real task from your line

Talk to us

C-Signal · Lab notes

Latest from the blog

All posts
Technology · Sep 3, 2026

Safety Is Architecture: Models Propose, Systems Decide

The model never directly controls actuators. CoreOS validates every command against four checks.
Research · Oct 15, 2026 · KO

Shipyard Physical AI briefing

Our regular public briefing on welding robots, digital shipyards and humanoids in production.
Trends · Sep 24, 2026

Plant Knowledge Agent vs Office Company Brain

A Core On view of on-prem plant knowledge agents versus office-centric company brains.
Trends · Sep 5, 2026

Agentic AI in Industry: What Public Reports Say

A briefing of public McKinsey, Accenture and IBM reports on industrial agentic AI adoption.

Company

Safety is architecture.
Data stays under customer control.

Headquarters
San Jose, California
Founded
2025
Founder
Paul Yu (Hyung Cik Yu)
Contact
paul.yu@coreon.ai

Bring a real task from your line.

Tell us what your robots need to do, and where the network, latency or data rules get in the way. We'll show you how the stack fits.

Talk to us Read the lab notes