URAI splits robot control between a programming agent and an execution agent. The programming agent creates and refines executable motion tools, while the execution agent selects and parameterizes those tools from observations and regains control after each complete motion.
Across five RoboDojo tasks and four frozen execution agents, aggregate success rose from 18.0% with direct fingertip control to 53.0%. Three agents also completed episodes 1.3 to 1.5 times faster while producing 1.5 to 1.7 times fewer execution-agent output tokens.
The key result is that reusable code does not have to replace model-level decision-making. Agents choosing tools after each call reached 56% success, compared with 24% for a program written in advance using the same tools.