LLM usecase support

[gitpqLee]

We need a LLM feature integrated into this project:

• Input: A ROS String Topic （see topic template） containing an instruction to AMR or manipulator arm.
• Output: A ROS topic (template)containing a navigation goal for mobile robot or a grasp pose for manipulator arm.

This repo should be added new code and logic to fulfill this Input->Inference->Output process: 1) Subscribe ROS topic of related audio common message. 2) Adopt LLM agent for the inference body. 3) Publish the corresponding topic(s).

See below picture as the pipeline:

LLM like Llama support can leverage OpenVINO-GenAI, please refer to OpenVINO™ GenAI Library

Development Scope:

• Wrap the audio to ROS topic.
• Use STT model convert audio to text [optional]
• Subscribe a ROS string topic and convert to input type of LLM agent.
• Prompt Intelligent agent to adapt usage (like navigation)
• Wrap the decode results to ROS topic, like PoseStamped.

You have the freedom to choose the intelligent agent.
Notes: Maybe here need a end-to-end LLM model like some VLM(Visual Language Model, for example VLM) models. Anyway, this is not easy to enable such a VLM model on Intel HW platform since there is no reference sample in OpenVINO currently.

[anurag1-dev]

Project_ROS2OV_Team2 is working on the issue