Overview
This research presents a language-model-based workflow designed for the generation, validation, and iterative correction of ABB RAPID robot programs. The system operates from natural language task descriptions, aiming to address the need for rapid reprogramming of industrial robots in flexible manufacturing environments, where product variants frequently change. The workflow integrates a dual-stream retrieval-augmented generation (RAG) pipeline for grounding code generation and a custom Model Context Protocol (MCP) server for connecting the language model to simulation software for automated feedback and correction.
Research Context
Flexible manufacturing necessitates the rapid reprogramming of industrial robots as product variants evolve. Traditional methods for robot programming can be time-consuming or require specialized expertise. The deployment of language models for code generation, particularly in domain-specific applications like industrial robotics, presents challenges related to generating accurate and executable programs. Ungrounded language models can produce domain-specific errors. Integrating verification and correction mechanisms, especially those leveraging simulation, is crucial for developing robust automated programming solutions.
Approach
The proposed workflow employs a multi-component architecture to generate and refine robot programs:
Dual-Stream Retrieval-Augmented Generation (RAG) Pipeline
This pipeline is designed to ground code generation. It utilizes two distinct streams to retrieve information, specifically:
- Verified technical documentation
- Production templates
The purpose of this dual-stream RAG is to reduce domain-specific errors that might arise from ungrounded language models during the code generation phase.
Model Context Protocol (MCP) Server
A custom MCP server facilitates the direct connection between the language-model client and ABB RobotStudio. This connection enables several automated functions:
- Automated code upload to the simulation environment.
- Execution of the generated robot program within the simulation.
- Provision of diagnostic feedback from the simulation back to the language model.
This protocol forms a closed-loop system, allowing the language model to iteratively correct programs based on simulation results.
Findings
The evaluation of the system involved several distinct components:
Retrieval Benchmark
A 30-query retrieval benchmark was conducted to assess the performance of the RAG pipeline. Specific results for this benchmark were not detailed in the abstract.
Scoped Code-Generation Checks
Code-generation checks were performed, but the abstract does not provide specific metrics or outcomes from these checks.
RobotStudio Case Studies
Case studies were executed within a simulated pick-and-place manufacturing cell using ABB RobotStudio. The simulation loop proved capable of exposing execution failures that static and semantic checks alone could not detect. These identified failures included:
- Suction release-height errors.
- Unreachable placement targets.
- Configuration-dependent recovery motions.
Overall, the results indicated that the integration of RAG and MCP successfully connected grounded code generation with executable feedback derived from industrial robot simulation software. While the system demonstrated its ability to generate and correct programs, it was observed to reduce, but not eliminate, the need for expert setup and final supervision.
Why This Matters
The ability to rapidly reprogram industrial robots is critical for achieving flexibility in modern manufacturing processes. This research addresses this need by providing a framework that automates aspects of robot program generation and validation. By leveraging natural language processing and simulation-based feedback, the workflow offers a method to produce functional robot programs more efficiently, potentially reducing the manual effort and specialized knowledge traditionally required. The identification of execution failures through simulation, which would otherwise be missed by static analysis, highlights the practical utility of integrating dynamic validation into the programming pipeline.