TypeGo OS runtime for robots cuts planning delays by 73%

A new paper from researchers at (affiliation not specified) introduces TypeGo, an operating-system-inspired runtime designed to address a fundamental problem in embodied AI: treating large language models (LLMs) as request/response oracles on the critical path is incompatible with real-time control and concurrent goals. TypeGo rethinks LLM-based planning by structuring it as asynchronous loops operating at multiple timescales that overlap with execution. The runtime manages the robot's physical body like an OS manages hardware: a Skill Kernel arbitrates typed physical subsystems among concurrent per-task processes, a scheduler preempts and resumes or replaces tasks by source, and speculative skill streaming hides LLM latency behind ongoing motion. A fast first-action path delivers visible feedback within one second. Users program behavior through natural language prescriptions that TypeGo dispatches to LLM-based planners or compiles into low-latency interrupt handlers.

The prototype runs on a Unitree Go2 quadruped, and preliminary results from their task suite are impressive: TypeGo cuts per-step delay by 50% over step-by-step planning and time-to-first-action by 73% over monolithic planning, while admitting concurrent tasks with low scheduling overhead. This design pushes embodied AI toward true real-time operation, where robots can react and adapt on the fly without waiting for a full LLM inference cycle. By abstracting the body as a set of typed resources and managing planning as a preemptible, streamable process, TypeGo offers a blueprint for building more responsive and capable autonomous systems.