Robots using an artificial intelligence brain repeatedly send growing conversation histories to a language model, while a separate vision-language-action system turns instructions into motions that must arrive on time. The brain’s key-value cache, a store of reusable information from earlier calculations, can fill up; when it cannot be kept, the system repeats work and slows down.
To measure this, the team recorded robot-task episodes, measured brain and action workloads, and replayed them on a shared model-serving system. They tested how the number of active robots affected speed, and whether admission control, meaning limiting which tasks may run at once, helped.
At thirty-two active robots, brain request throughput fell to about one-quarter of its earlier level. Limiting active episodes brought throughput back and allowed more complete robot programs to finish. This matters because fleet capacity depends not just on motion deadlines but also on keeping reusable brain state available.