Multi-Agent RAG for Autonomous Vehicles Using Decentralized Knowledge Graph on Blockchain

Autonomous agricultural vehicles operate in dynamic environments where isolated learning and centralized coordination limit scalability and adaptability. To address this, the paper proposes a decentralized multi-agent Retrieval Augmented Generation (RAG) framework for autonomous farming vehicles that allows for collaborative real-time decision making via a blockchain-based distributed knowledge graph. Each vehicle node functions as an agentic entity that retrieves the validated field knowledge, such as current soil condition patterns, crop stress indicators, and the historical traversal results along with their outcomes from the shared knowledge graph, which are then integrated with the local sensory observations to then generate dynamic context-aware operational decisions. Blockchain provides a tamper-proof layer for knowledge integrity and authenticity, which ensures that the experiential updates are validated using lightweight consensus mechanisms before they are circulated, thereby preventing erroneous or malicious knowledge propagation. As opposed to existing blockchain-based agricultural solutions that focus primarily on data logging to prevent tampering and misuse, the proposed framework integrates blockchain directly into the agent's reasoning and learning loop. Experimental evaluation on a synthetically generated dataset of 5,000 interaction instances demonstrates that the proposed framework achieves a task success rate of 88.6%, compared to 81.0% for decentralized multi-agent baselines and 79.5% for centralized approaches. The framework further reduces latency by up to 28% while improving knowledge utilization by 10-19% and significantly lowering error propagation by up to 66%, indicating more stable and reliable decision-making, representing an improvement of approximately 9-11% over baseline methods. The results indicate that decentralized knowledge-driven reasoning can enhance the robustness and long-term learning in autonomous agricultural vehicle networks.