Understanding and action are not two departments. They are one loop.

The knowledge a robot needs is indexical — it is about here, now, this. Not “cups are graspable” but this cup, at this pose, reachable by this arm whose second joint runs four degrees hot. Most machine intelligence today manipulates symbols that refer to nothing in particular. Robot software has no such luxury: its symbols cash out at sensors and motors, or they don’t cash out at all.

So we build perception and state estimation that keep the connection between symbol and situation intact — from raw phenomena to the planner and back — instead of laundering it away through a benchmark.

“Roughly right” compounds into completely wrong within a few actions. Contact, friction, deformation, occlusion — the physical world charges interest on every approximation. Understanding reality at the granularity where robots live means modeling it at the granularity where physics lives, and being honest about the residual.

That honesty is an engineering artifact, not a virtue: calibrated uncertainty, explicit failure envelopes, models that know when they are out of their depth before the gripper does.

The world does not replay. Everything reproducible about a robot’s behavior — training, search, regression, ablation — belongs in simulation and replay, where an experiment can be run twice. Everything that matters ultimately happens once, in reality, where it can’t.

We treat the gap between those two as a first‑class quantity: measured continuously, narrowed deliberately, and never papered over with a demo reel. Environments and evaluation are core infrastructure here, not an afterthought bolted on before a launch.

Perceive, understand, decide, act — and then perceive what the action did. The second perception is the one most systems skip, and it is the one that turns a pipeline into intelligence. Acting on the world and being corrected by it is the only supervision signal reality offers for free, at scale, forever.

The ring we use as a mark is drawn in one continuous stroke. Understanding flows into action, action back into understanding. Break the circle anywhere and both halves die. That is the thesis, and everything we ship is judged by whether it strengthens the circle.