Splitting one big instruction into two smaller ones makes robots better planners.
Researchers building general purpose service robots for the RoboCup@Home competition ran into a familiar problem: cramming a full command, world context, and action options into a single prompt causes large language models to lose track of instructions buried in the middle. Their fix separates the job into two specialized stages, one that classifies what kind of command it is and another that generates the action sequence, cutting the prompt length fed to each stage by about 45 percent. Tested against 100 randomly generated GPSR commands across three models, including both open-source and frontier cloud systems, the two-stage approach beat the single-prompt method every time, with local open-source models gaining as much as 37 percentage points in planning accuracy. On Toyota's HSR robot, though, only 6 of 10 tasks that were planned successfully actually got completed.
That gap matters more than the planning numbers do. It is a reminder that language models are good at describing what a robot should do and much worse at guaranteeing a robot's arms, sensors, and grippers can pull it off. Most of the AI industry's chaining and agent-routing techniques were built to fix reasoning failures, not physical ones, so this result draws a clear line between the two kinds of problems.
Home robots keep getting smarter conversationalists and staying clumsy hands. Better prompting will not fix a gripper that misses a cup.