When a line stops, that is yield, not latency. The moment control requires a cloud hop, that loop is no longer the factory’s. Industrial AI runs on the edge.
Reaction and Planning, Same Box
CoreVLA is the seeing-and-grasping side. See the object, understand the context, act. Picking, grasping, navigation. System 1.
CoreLM handles planning and explanation. “Why did this fail” “What’s next”. Domain LLM. System 2.
They live on the same device for a simple reason. You need to question the action where it happened, immediately. The moment you introduce a cloud round-trip, the loop breaks.
The Cloud Alone Does Not Run Factories
Cloud AI solved many problems. But factories are different.
Time: Network RTT prevents real-time control.
Reliability: Network drops, robot stops. Factories run 24/7.
Sovereignty: Sending production data out raises contractual and regulatory issues.
Cost: Fleet-scale API calls mean linear cost scaling.
We run both models on NVIDIA Jetson Thor-class hardware (CoreEdge). This is principle, not compromise.
Models Propose, Systems Grant
When CoreVLA proposes an action, CoreLM enriches context if needed. Every output passes through CoreOS. The deterministic safety layer. Between model output and actuators.
Models propose. Systems grant permission.
The next post covers the safety architecture. Why architecture is responsible for safety, and how the data flywheel turns.
Core On Dynamics, San Jose, California.