RAG and MCP Generate, Correct Robot Programs

The demand for flexible manufacturing necessitates rapid reprogramming of industrial robots to adapt to changing product variants. This research presents an innovative workflow that leverages language models to generate, validate, and iteratively correct ABB RAPID robot programs directly from natural language task descriptions. The system employs a dual-stream Retrieval-Augmented Generation (RAG) pipeline. This pipeline grounds the code generation process in verified technical documentation and production templates, significantly reducing the domain-specific errors that ungrounded language models often produce. A crucial component is a custom Model Context Protocol (MCP) server, which establishes a direct connection between the language model client and ABB RobotStudio. This connection enables automated code uploads, simulation execution, and diagnostic feedback. Evaluation involved a retrieval benchmark, code-generation checks, and RobotStudio case studies within a simulated pick-and-place manufacturing cell. The simulation loop proved vital in uncovering execution failures that static and semantic checks alone could not detect, such as suction release-height errors or unreachable targets. The results demonstrate how RAG and MCP effectively link grounded code generation with executable feedback from industrial robot simulation software, streamlining robot programming while still requiring expert setup and final supervision.