ROSGPT_Vision Download - ROSGPT_Vision Source code download

ROSGPT_Vision

AI Source Code

1.0.0

Download

ROSGPT_Vision: Commanding Robots Using Only Language Models' Prompts

Bilel Benjdira, Anis Koubaa and Anas M. Ali

Robotics and Internet of Things Lab (RIOTU Lab), Prince Sultan University, Saudi Arabia

Inspired by ROSGPT. Both projects aim to bridge the gap between robotics, natural language understanding, and image analysis.

Collaborators who want to participate in this project, are very welcome.

ROSGPT_Vision is a new robotic framework dsigned to command robots using only two prompts:
- a Visual Prompt (for visual semantic features), and
- an LLM Prompt (to regulate robotic reactions).
It is based on a new robotic design pattern: Prompting Robotic Modalities (PRM).
ROSGPT_Vision is used to develop CarMate, a robotic application for monitoring driver distractions and providing real-time vocal notifications. It showcases cost-effective development.
We demonstrated how to optimize the prompting strategies to improve the application.
LangChain framework is used by to easily customize prompts.
More details are described in the academic paper "ROSGPT_Vision: Commanding Robots using only Language Models' Prompts".

Video Demo

An illustrative video demonstration of ROSGPT_Vision is provided: ROSGPT Video Demonstration

Overview

ROSGPT_Vision offers a unified platform that allows robots to perceive, interpret, and interact with visual data through natural language. The framework leverages state-of-the-art language models, including LLAVA, MiniGPT-4, and Caption-Anything, to facilitate advanced reasoning about image data. LangChain is used for easy customization of the prompts. The provided implementation includes the CarMate application, a driver monitoring and assistance system designed to ensure safe and efficient driving experiences.

ROSGPT_Vision diagram

Prompting Robotic Modalities (PRM) Design Pattern

A new design approach emphasizing modular and individualized sensory queries.
Uses specific Modality Language Models (MLM) for textual interpretations of inputs, like the Vision Language Model (VLM) for visual data.
Ensures precise data collection by treating each sensory input separately.
Task Modality's Role: Serves as the central coordinator, synthesizing data from various modalities.

** for more information go to

CarMate Application

CarMate is a complete application for monitoring driver behavior which was developed just by setting two prompts in the YAML file. It automatically analyses the input video using the Visual prompt, analyses what should be done using the LLM prompt, and gives an instant alert to the driver when needed.

These are the prompts used to develop the application, without needing extra code:

The Visual prompt:

Visual prompt: "Describe the driver’s current level of focus 
on driving based on the visual cues, Answer with one short sentence."

The LLM prompt:

LLM prompt:"Consider the following ontology: You must write your Reply 
with one short sentence. Behave as a carmate that surveys the driver 
and gives him advice and instruction to drive safely. You will be given 
human language prompts describing an image. Your task is to provide 
	appropriate instructions to the driver based on the description."

We can see three examples of scenarios, got during the driving:

Scenario 1: The driver is using phone

We can see in the top box the description generated by the image semantics module for the input image using the Visual prompt. Meanwhile, the second box generates the alert that should be given to the driver using the LLM prompt.

Scenario 2: The driver is taking pictures

Scenario 3: The driver is drinking

Installation

To use ROSGPT_Vision, follow these steps:

1. Prepare the code and the environment

Git clone our repository, creating a python environment and ativate it via the following command

  git clone https://github.com/bilel-bj/ROSGPT_Vision.git
  cd ROSGPT_Vision
  git clone https://github.com/Vision-CAIR/MiniGPT-4.git
  git clone https://github.com/haotian-liu/LLaVA.git
  conda env create -f environment.yml
  conda activate ROSGPT_Vision

2. Install the required dependencies

You can run image_semantics.py by install all required dependencies from LLAVA, MiniGPT-4 and Caption-Anything.
Ensure the installation of all requisite dependencies for ROS2.

Usage

To regulate all parameters associated with ROSGPT_Vision, modifications can be made within the corresponding .yaml file.

The YAML contains 6 main sections of configurations parameters:

Task_name: This field specifies the name of the task that the ROS system is configured to perform.
ROSGPT_Vision_Camera_Node: This section contains the configuration for the ROSGPT_Vision_Camera_Node.
Image_Description_Method: This field specifies the method used by the node to generate descriptions from images. It can be one of the currently developed methods: MiniGPT4, LLaVA, or SAM. The configurations needed for everyone of them is put separately at the end of this file.
Vision_prompt: This field specifies the prompt used to guide the image description process.
Output_video: This field specifies the path or the name of where to save the output video file.
GPT_Consultation_Node: This section contains the configuration for the GPT_Consultation_Node.
- llm_prompt: This field specifies the prompt used to guide the language model.
- GPT_temperature: This field specifies the temperature parameter for the GPT model, which controls the randomness of the model's output.
MiniGPT4_parameters: This section contains the configuration for the MiniGPT4 model. It should be clearly set if the model is used in this task, otherwise it could be empty.
- configuration: This field specifies the path for the configuration file of MiniGPT4.
- temperature_miniGPT4: This field specifies the temperature parameter for the MiniGPT4 model.
llava_parameters: This section contains the configuration for the llavA model (if used).
- temperature_llavA: This field specifies the temperature parameter for the llavA model.
SAM_parameters: This section contains the configuration for the SAM model.
- weights_SAM: This field specifies the weights used by the SAM model.

Run in Terminal local machine

run first terminal :

        colcon build --packages-select rosgpt_vision
		    source install/setup.bash
		    python3 src/rosgpt_vision/rosgpt_vision/rosgpt_vision_node_web_cam.py
		    python3 src/rosgpt_vision/rosgpt_vision/ROSGPT_Vision_Camera_Node.py /home/anas/ros2_ws/src/rosgpt_vision/rosgpt_vision/cfg/driver_phone_usage.yaml

run second terminal:

        colcon build --packages-select rosgpt_vision 
		    source install/setup.bash
		    python3 src/rosgpt_vision/rosgpt_vision/ROSGPT_Vision_GPT_Consultation_Node.py /home/anas/ros2_ws/src/rosgpt_vision/rosgpt_vision/cfg/driver_phone_usage.yaml

run third terminal:

bash ros2 topic echo /Image_Description

run fourth terminal:

bash ros2 topic echo /GPT_Consultation

Citation

@misc{benjdira2023rosgptvision,
  title={ROSGPT_Vision: Commanding Robots Using Only Language Models' Prompts}, 
  author={Bilel Benjdira and Anis Koubaa and Anas M. Ali},
  year={2023},
  eprint={2308.11236},
  archivePrefix={arXiv},
  primaryClass={cs.RO}
  }

License

This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License. You are free to use, share, and adapt this material for non-commercial purposes, as long as you provide attribution to the original author(s) and the source.