Perception
Use cameras, sensors and multimodal models to understand objects, spaces, conditions and events.
Intelligence in the physical world
Physical AI: When Intelligence Leaves the Screen
Connect AI to robots, sensors, simulation and real environments while keeping action boundaries, human authority and infrastructure requirements explicit.
From digital intelligence to physical action
When intelligence leaves the screen
Generative AI produces information. Agentic AI can perform bounded digital actions. Physical AI connects models to sensors, spatial context, machines, robots and real environments.
That transition changes the governance problem. A model error that produces poor text can create misinformation. A system error that controls a machine can create an incorrect physical action.
Physical AI requires performance engineering, simulation, safety boundaries and operational governance to meet in the same architecture.
Physical AI stack
Intelligence that can perceive, reason and act
AFA is developing practical capability around the full environment required for governed Physical AI.
Simulation
Test machines, workflows, digital twins and operating scenarios before physical deployment.
Robotics
Connect intelligence to manipulators, mobile systems and specialized machines for bounded tasks.
Spatial intelligence
Reason about geometry, location, environments and interaction in three-dimensional space.
Compute placement
Coordinate central accelerated computing with edge inference where latency, bandwidth or resilience require it.
Operational sovereignty
Define what the system may do, which information it can use, what requires human approval and how actions are recorded.
AFA Reference Lab
A place to test Physical AI before real-world scale
The proposed Toronto-based Sovereign AI & Physical AI Reference Lab is intended to combine Dell Technologies and NVIDIA accelerated-computing infrastructure with model evaluation, simulation, agent testing and robotics development.
The objective is to build practical evidence around how intelligent physical systems should be architected, evaluated, coordinated and governed—not to present untested autonomy as a finished product.
Operational lifecycle
Perceive → simulate → validate → act → improve
Physical AI should advance by evidence and bounded authority.
Define the task
Specify the physical objective, environment, constraints and human authority.
Simulate
Test models, perception and behaviour in digital or controlled environments.
Validate
Measure performance, failure modes, latency, recovery and governance boundaries.
Deploy bounded
Introduce physical capability with explicit permissions, fallback and oversight.
Learn safely
Use operational evidence to improve without silently expanding authority.
Continue the journey
Related pathways
Practical perspective
From understanding to practical action
Atkinson connects people, policy, infrastructure and implementation so each initiative can move forward with clarity, accountability and purpose.
Bring a Physical AI Problem to the Lab
Start with the task, environment, performance requirements and authority model.
Request a Physical AI Briefing