5.4.2 LLM 模型聊天

功能简介

本模块基于本地部署的大语言模型（LLM, Large Language Model），实现自然语言处理对话能力，支持非流式服务与流式 Action 两种交互模式。

本模块可广泛应用于以下场景：

• 人机自然语言交互
• 智能问答与知识查询
• 多轮对话状态管理
• 指令意图理解与生成

环境准备

安装系统依赖

sudo apt update
sudo apt install -y libopenblas-dev \
	portaudio19-dev \
	python3-dev \
	ffmpeg \
	python3-spacemit-ort \
	libcjson-dev \
	libasound2-dev \
	python3-pip \
	python3-venv

运行环境配置与依赖安装

# 安装 LLM 模型运行环境（含模型下载与配置）
bash /opt/bros/humble/share/br_chat/llm_setup.sh

# 创建并激活 Python 虚拟环境
pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
pip config set global.extra-index-url https://git.spacemit.com/api/v4/projects/33/packages/pypi/simple

python3 -m venv ~/ai_env
source ~/ai_env/bin/activate

# 安装 Python 依赖
pip install -r /opt/bros/humble/share/br_chat/requirements.txt

非流式聊天服务接口

导入环境

source /opt/bros/humble/setup.bash
source ~/ai_env/bin/activate
export PYTHONPATH=~/ai_env/lib/python3.12/site-packages/:$PYTHONPATH

启动服务端

ros2 launch br_chat chat_service.launch.py

默认服务名为：chat_service，服务类型为 jobot_ai_msgs/srv/LLMChat

客户端调用示例

新建 llm_client.py 文件：

import rclpy
from rclpy.node import Node
from jobot_ai_msgs.srv import LLMChat

class LLMClientNode(Node):
    def __init__(self):
        super().__init__('llm_client')
        self.cli = self.create_client(LLMChat, 'chat_service')
        while not self.cli.wait_for_service(timeout_sec=1.0):
            self.get_logger().info('等待 LLM 服务中...')
        self.req = LLMChat.Request()

    def send_request(self, prompt_text):
        self.req.prompt = prompt_text
        future = self.cli.call_async(self.req)
        return future

def main(args=None):
    rclpy.init(args=args)
    node = LLMClientNode()
    future = node.send_request("你好，告诉我什么是 ROS2？")
    rclpy.spin_until_future_complete(node, future)

    if future.result() is not None:
        print("模型回复:", future.result().response)
    else:
        node.get_logger().error('调用服务失败')

    node.destroy_node()
    rclpy.shutdown()

if __name__ == '__main__':
    main()

执行：

python3 llm_client.py

输出示例：

模型回复: Hello! ROS (Robot Operating System) is a set of development tools and runtime environment developed by the ROS Core Team. It provides complete system for building real-time autonomous robots.

It consists of three main components: **Core**, **Dynamic** and **Toolbox**.
- The core includes fundamental systems like sensor abstraction, command line interface, communication protocol handling, etc.
- Dynamic component is responsible to control robot execution based on user input
- Toolbox provides a set of tools for building custom modules or applications with the help of ROS.

With ROS2 you can create robots that have self-awareness and autonomous decision-making capabilities. It also has libraries like **Bullet**, which are used by ROS to simulate real-time systems, etc.
Is there anything else I should know about it?

流式聊天 Action 接口

支持实时响应输出的 LLM 聊天动作，适用于语音或多轮交互响应流式推送场景。

导入环境

source /opt/bros/humble/setup.bash
source ~/ai_env/bin/activate
export PYTHONPATH=~/ai_env/lib/python3.12/site-packages/:$PYTHONPATH

启动 Action 服务端

ros2 launch br_chat chat_action.launch.py

默认 Action 名称为 chat_action，类型为 jobot_ai_msgs/action/LLMChatAction。

客户端调用示例

新建 llm_client_action.py 文件：

import rclpy
from rclpy.node import Node
from rclpy.action import ActionClient
from jobot_ai_msgs.action import LLMChatAction

class LLMChatClient(Node):
    def __init__(self):
        super().__init__('llm_chat_client')
        self._action_client = ActionClient(self, LLMChatAction, 'chat_action')

    def send_prompt(self, prompt):
        goal_msg = LLMChatAction.Goal()
        goal_msg.prompt = prompt

        self._action_client.wait_for_server()

        self._send_goal_future = self._action_client.send_goal_async(
            goal_msg,
            feedback_callback=self.feedback_callback
        )
        self._send_goal_future.add_done_callback(self.goal_response_callback)

    def feedback_callback(self, feedback_msg):
        feedback = feedback_msg.feedback
        print(f'{feedback.partial}', end='', flush=True)

    def goal_response_callback(self, future):
        goal_handle = future.result()
        if not goal_handle.accepted:
            self.get_logger().info('目标被拒绝')
            return

        self.get_logger().info('目标已被接受')
        self._get_result_future = goal_handle.get_result_async()
        self._get_result_future.add_done_callback(self.result_callback)

    def result_callback(self, future):
        result = future.result().result
        print(f'\n\n[Result] {result.response}')
        rclpy.shutdown()

def main(args=None):
    rclpy.init(args=args)
    client = LLMChatClient()

    import sys
    if len(sys.argv) > 1:
        prompt = ' '.join(sys.argv[1:])
    else:
        prompt = "你好，请介绍一下你自己"

    client.send_prompt(prompt)
    rclpy.spin(client)

if __name__ == '__main__':
    main()

执行：

python3 llm_client_action.py

终端输出示例：

[INFO] [1748593600.839893117] [llm_chat_client]: 目标已被接受
我是由阿里云开发的大规模多语言模型，可以提供各种类型的信息和解答您的问题，如新闻、财经资讯、科技知识等。如果您有任何具体的问题或需要帮助的事项，请随时告诉我！

[Result] 我是由阿里云开发的大规模多语言模型，可以提供各种类型的信息和解答您的问题，如新闻、财经资讯、科技知识等。如果您有任何具体的问题或需要帮助的事项，请随时告诉我！

总结对比

模式	接口类型	优点	场景
非流式服务	Service	简单易用，响应完整	问答类应用、任务调用
流式输出	Action	实时反馈，适配语音交互与长文本生成	多轮对话、语音播报等流式交互