Deployments Are Hard

Physical AI deployments are hard. Why? Because robots aren't reliable enough. Why? Because models aren't good enough. Why? Because there isn't enough data. Why? Because there aren't enough deployments (and we've only scratched the surface on all the data that we need).

For deployments to work, the solution is a hybrid deployment approach: start with teleoperation until robots can operate autonomously, then transition to human-in-the-loop interventions and DAgger to collect more data, specifically in failure & recovery scenarios, improving reliability and teaching robots how to recover from failures as deployments scale.

There's an entire layer of infrastructure that needs to be built. That includes insurance, hardware maintenance, network systems, etc.

The process of taking a robot from the lab to a customer site and getting it operational, eventually reaching autonomy, needs to be streamlined.

Deploying a robot should be as easy and frictionless as deploying software.