The Blueprint for Physical AI

The deployment gap
Today's demos deliver about 90%. The industry runs at 99.9%. The missing 10% is not a rounding error. It is a stopped supply chain, a broken machine, an injured worker.
Physical common sense
Foundation models need engineering to deliver, not just more data and training. What agentic coding did for software, robots will do for physical work. The parallels:
| Agentic coding | Physical AI |
|---|---|
| Tools | Robot skills |
| Security | Safety |
| Validation | Simulation |
The question is how to add new reasoning without breaking what already works.
Engineer for safety and adoption
Safety assumes zero trust. A system must be safe even with the dumbest AI in charge, enforced through contracts, envelopes and guardrails. On top of that sit robot skills and orchestration: perception, locomotion and manipulation as modules an agent can call. The open question is how to tell whether AI makes the work more productive and keeps it safe.
Be serious about simulation
Test operational behaviour at scale before anything touches the floor. Simulation is used to train policies and models and to prototype and fail fast. The shift in emphasis is toward the third use: testing operations at scale, with metrics. That needs a simulation toolbox built on interoperability and standards such as OpenUSD.
Small wins
Start with AI that is auxiliary and additive: each use adds value without replacing core processes.
- Inspection: compliance and safety, by general rules and site-specific ones.
- Debugging: faster root cause, less downtime.
- "Stuck" resolution: reason and rescue without human intervention.
The blueprint
- Compute: inference-ready.
- Simulation: cover the whole cycle.
- Robot skills: a modular framework, not a monolith.
- Adoption: small wins first.
Time to build
AI doesn't have to do everything to be useful.
Learn from the success of agentic coding, and engineer for safety and adoption.
Related work: the ROS Simulation Interfaces, REP 158 for OpenUSD robot assets, and RAI, an agent framework for embodied AI.