Giving robots a usable model of the world
A robot can detect a cup. Whether it knows a cup can hold liquid, tips over, and belongs on a table — that’s a different system entirely.
Perception is not understanding
Modern perception is excellent at labels and poor at consequences. Physical AI is the attempt to close that gap: to give an agent structured knowledge it can reason over, not just a stream of detections.
Ontologies as the substrate
Frameworks like KnowRob represent objects, affordances, and actions as a queryable knowledge base. The robot asks questions — what can I do with this? what happens if I tip it? — and gets answers grounded in an explicit model rather than a black box.
It’s unfashionable next to end-to-end learning, but for anything that has to act safely in the real world, an inspectable model of why is worth a lot. Ongoing notes live in knowrob_OPAL.