1. A foundation model has to drive the hand

   Scripted motion can't absorb the variability of real work. Getting human-level dexterity means a general robot foundation model in the loop, not a fixed program per part.

2. That model needs data nobody has

   A model is only as good as what it's post-trained on. Manipulation needs a large corpus of high-precision demonstrations, including the contact and tactile force information that vision alone can never capture.

3. Collecting it means wearing the rig all shift

   So Aero UMI is ergonomic and lightweight before it is anything else. If an operator can't comfortably wear the device for a full shift, the data pipeline breaks before it starts.

4. And the hand has to survive deployment

   Strong, fast, durable, reliable. All that data is worth nothing if the hardware it trains can't hold up through millions of cycles on a real production line.

Which is why we build both. One system collects the data and runs the deployment — same morphology, same tactile sensing, zero embodiment gap between the human demonstration and the robot that has to repeat it.

The approach is already proven. UMI‑style handheld data collection has produced strikingly capable manipulation policies on simple grippers. For a high‑DOF dexterous hand — where the action space is an order of magnitude larger and contact force carries most of the signal — that pipeline becomes more essential, not less.