ACE Robotics has described ACE-CaP-0, a system in which an AI agent writes and then freezes robot control code rather than issuing commands step by step.
The distinction matters for reliability. A model that plans every action at runtime inherits the latency and variability of inference; a model that generates a validated program once, then executes deterministically, behaves like ordinary industrial software.
The authors report 221 of 240 successful real-robot trials, or 92.1 per cent, from their own testing. That is a credible figure for structured manipulation tasks and a reminder that the remaining 8 per cent is where deployment cost lives.
Reusability is the commercial hook: generated programs can be inspected, versioned and reused, which fits the way manufacturing engineers already work.
The open question is coverage - how often the agent can produce a correct program for a novel task, and what happens when it cannot.
