When Large Language Models Take Over Automotive CAN Bus: In-Depth Analysis of the LLM-CAN-Interface Intelligent Safety Controller

This article introduces the LLM-CAN-Interface intelligent safety control system, which integrates a locally deployed Llama 3.1 large language model with the automotive CAN bus to analyze vehicle telemetry data (such as engine speed, vehicle speed, coolant temperature, tire pressure, etc.) in real time, determine the vehicle's safety status, and autonomously execute intervention commands. Addressing the increasing complexity of vehicle electronic systems in the autonomous driving era and the lack of flexibility in traditional rule-based systems, this project pioneers a new paradigm of AI-driven intelligent vehicle safety control.

With the rapid development of autonomous driving and intelligent connected vehicles, the complexity of vehicle electronic systems has grown exponentially. Traditional rule-based safety control systems lack flexibility and adaptability in complex and changing driving environments. The core innovation of the LLM-CAN-Interface project lies in: using a locally deployed Llama 3.1 (8B parameter version) to analyze vehicle data in real time, judge safety boundaries through natural language reasoning, and send intervention commands via the CAN bus when necessary.

Hardware Architecture: Uses the WeAct Studio USB-to-CAN/CANFD module (based on the STM32G4 chipset, MCP2515 emulates the SLCAN protocol, supports 500kbps transmission); AI inference relies on the NVIDIA GeForce RTX3060 Laptop GPU (6GB VRAM), ensuring an inference latency of 200-500ms.

Software Stack: Docker containerized deployment of the Ollama inference service (GPU passthrough via NVIDIA Container Toolkit); Python 3.10+ as the main development language, with serial port libraries for hardware interaction.

System inference process: CAN bus collects sensor data → formatted into structured natural language and input to Llama 3.1 → model infers based on safety knowledge base → outputs natural language decision → converted into CAN commands. For example: input "Engine speed 3500 RPM, vehicle speed 150 km/h, coolant temperature 115°C, tire pressure 2.0 bar", the model judges overheating and needs to limit speed, converting to a speed limit command with CAN ID 0x316. Compared to traditional hard-coded rules, this method can handle complex and ambiguous scenarios, has strong context understanding capabilities, and can quickly adapt to new strategies by updating the model's knowledge base.

Case 1: Engine Overheating Protection When the coolant temperature reaches 115°C (exceeding the 100°C threshold), the AI judges that speed needs to be limited, sends the CAN ID 0x316 command t31680BB8000000000000, setting the maximum speed to 3000 RPM.

Case 2: Tire Pressure Abnormality Handling Tire pressures are 1.7/1.5/1.4/1.5 bar (minimum 1.4 bar is dangerous), the AI judges that speed needs to be limited, sends the CAN ID 0x320 command t32083400000000000000, limiting the vehicle speed to 52 km/h.

Core code modules:

1. main.py: Handles serial communication with the WeAct USB-CAN module and parses raw CAN frames;
2. ai_engine.py: Encapsulates prompt engineering and Ollama API communication, converting data into a format understandable by AI;
3. message_send.py: Formats Lawicel/SLCAN commands and sends them.

Deployment: Define container configurations (GPU resource allocation, port mapping, environment variables) via docker-compose.yml to enable one-click startup of the AI service.

This project is only authorized for educational, research, or Hardware-in-the-Loop (HIL) simulation testing; commercial use or deployment on real vehicles is strictly prohibited. Technically, the 200-500ms inference latency is insufficient for millisecond-level response scenarios (such as emergency braking). In the future, LLM can be used as a high-level decision-making engine, while critical safety functions are executed by traditional deterministic systems.

LLM-CAN-Interface reveals the trend of large language models penetrating into physical control systems. The integration of "AI + Embedded" opens up new possibilities for intelligent vehicles, industrial automation, and other fields. In the future, improvements in edge computing capabilities and optimization of model inference efficiency will promote the implementation of more similar systems. At the same time, software-hardware collaboration is needed to ensure the safe, real-time, and reliable operation of the system in critical scenarios.