Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLCoach: Generating Robot Soccer Plans using Multi-Role Large Language Models

Michele Brienza, Emanuele Musumeci, Vincenzo Suriani, Daniele Affinita, Andrea Pennisi, Daniele Nardi, Domenico Daniele Bloisi

TL;DR

This paper tackles autonomous robot soccer planning in dynamic RoboCup SPL environments by proposing LLCoach, a multi-role LLM/VLM-driven pipeline that generates and refines executable multi-agent plans. The approach combines Retrieval-Augmented Generation with a Semantic Action Database, a Visual Language Model coach (via GPT-4V) and the MARIO visual analysis tool to ground plans to discrete waypoints and STRIPS-like actions. Plan refinement and synchronization ensure multi-agent coordination, yielding structured, groundable plans suitable for NAO robots. In simulation against a human-coded baseline, LLCoach achieves substantially higher success rates and faster scoring, demonstrating the viability of generative AI for embodied multi-robot planning in competitive settings.

Abstract

The deployment of robots into human scenarios necessitates advanced planning strategies, particularly when we ask robots to operate in dynamic, unstructured environments. RoboCup offers the chance to deploy robots in one of those scenarios, a human-shaped game represented by a soccer match. In such scenarios, robots must operate using predefined behaviors that can fail in unpredictable conditions. This paper introduces a novel application of Large Language Models (LLMs) to address the challenge of generating actionable plans in such settings, specifically within the context of the RoboCup Standard Platform League (SPL) competitions where robots are required to autonomously execute soccer strategies that emerge from the interactions of individual agents. In particular, we propose a multi-role approach leveraging the capabilities of LLMs to generate and refine plans for a robotic soccer team. The potential of the proposed method is demonstrated through an experimental evaluation,carried out simulating multiple matches where robots with AI-generated plans play against robots running human-built code.

LLCoach: Generating Robot Soccer Plans using Multi-Role Large Language Models

TL;DR

This paper tackles autonomous robot soccer planning in dynamic RoboCup SPL environments by proposing LLCoach, a multi-role LLM/VLM-driven pipeline that generates and refines executable multi-agent plans. The approach combines Retrieval-Augmented Generation with a Semantic Action Database, a Visual Language Model coach (via GPT-4V) and the MARIO visual analysis tool to ground plans to discrete waypoints and STRIPS-like actions. Plan refinement and synchronization ensure multi-agent coordination, yielding structured, groundable plans suitable for NAO robots. In simulation against a human-coded baseline, LLCoach achieves substantially higher success rates and faster scoring, demonstrating the viability of generative AI for embodied multi-robot planning in competitive settings.

Abstract

The deployment of robots into human scenarios necessitates advanced planning strategies, particularly when we ask robots to operate in dynamic, unstructured environments. RoboCup offers the chance to deploy robots in one of those scenarios, a human-shaped game represented by a soccer match. In such scenarios, robots must operate using predefined behaviors that can fail in unpredictable conditions. This paper introduces a novel application of Large Language Models (LLMs) to address the challenge of generating actionable plans in such settings, specifically within the context of the RoboCup Standard Platform League (SPL) competitions where robots are required to autonomously execute soccer strategies that emerge from the interactions of individual agents. In particular, we propose a multi-role approach leveraging the capabilities of LLMs to generate and refine plans for a robotic soccer team. The potential of the proposed method is demonstrated through an experimental evaluation,carried out simulating multiple matches where robots with AI-generated plans play against robots running human-built code.

Paper Structure

This paper contains 18 sections, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Approach
  4. Action Retrieval
  5. Coach VLM
  6. Visual Analysis.
  7. VLM module.
  8. Role Retrieval.
  9. Plan Refinement
  10. Plan Grounding.
  11. Plan Synchronizer.
  12. Plan Execution
  13. Experimental Results
  14. Prompt Engineering.
  15. Setup.
  16. ...and 3 more sections

Figures (3)

  • Figure 1: The LLCoach architecture includes high-level plan generation by the VLM coach, which is then refined to a low-level plan, executable in robot matches.
  • Figure 2: The pipeline is subdivided into an offline component and an online one, executed in real-time. The offline component collects plans by passing video frames to the coach VLM, and refines them using a multi-role LLM pipeline, using only actions obtained by RAG. The online component retrieves and executes the most fitting plan according to the world model, shared between robots.
  • Figure 3: MARIO visual tool, transforming robot pose (original (a), simulation (b)).