The entire pipeline was tested in Ubuntu 22.04 with ROS2 Humble Hawksbill and Gazebo Classic. The zip file contains all the necessary code to run the pipeline. Please note that folders custom_msgs, sim_gazebo, turtlebot3_rl inside the zip are ros2 packages. All others are ordinary directories.
We have used Groq AI Inference Engine. So the LLM requires the API key from Groq to run. Please do login in Groq, get API keys and put in the below file
tool_caller.py in llm/LLM/LLM, line 18 and 35. Put your API Key as a string in Groq(api_key=<your_API_key>)
There is a gazebo_models.zip file. This contains the models and obstacles in gazebo. Please extract them to .gazebo/models (hidden folder) in your home directory.
A few videos are in the Videos folder which can help one understand how to run the pipeline. However, the steps have been mentioned in detail below.
mkdir -p swarm_ws/src
cd swarm_ws/src
Unzip the files in the src directly. Then, colcon build the workspace by going one step back
cd ..
colcon build
There is a requirements.txt file including all the necessary python libraries. To install do the following
cd src
pip3 install -r requirements.txt
This file contains the robot positions, goal positions, RL algorithm hyperparameters, and some import variables for a 10x10 environment
This is the main launch file that spawns 4 robots, 3 dynamic obstacles, and 4 targets in a 10x10 environment. The associated positions of the robot are read from params.yaml
This file contains the robot positions, goal positions, RL algorithm hyperparameters, and some import variables for a 15x15 environment
This is the main launch file that spawns 5 robots, 3 dynamic obstacles, and 5 targets in a 15x15 cluttered environment. The associated positions of the robot are read from params_large.yaml
This file holds the robot coordinates, target coordinates, status of the robot, and the assigned target to a robot in the 10x10 environment. This file is extensively used by the PyGame Visualizer, LLM Chatbot and RL Agent to update status of the robot.
This file holds the robot coordinates, target coordinates, status of the robot, and the assigned target to a robot in the 15x15 environment.
Terminal 1: LLM Main Script (Make sure you point to src)
cd swarm_ws/src
python3 python3 llm/LLM/LLM/app.py
Terminal 2: Streamlit Interface (Make sure you point to src)
streamlit run llm/LLM/LLM/chat_ui.py
Use the streamlit UI to assign robot the targets. An example prompt is given below
assign robot_1 to service_1, robot_2 to charging_2, robot_3 to vending_machine, robot_4 to charging_1
Terminal 3: Spawn 4 bots launch file (Make sure you point to swarm_ws)
cd ..
source install/setup.bash
ros2 launch sim_gazebo spawn_4_bots.launch.py
Terminal 4: Obstacle Detection Script (Uses laser scan to visualize object locations in RViZ) (Make sure you point to swarm_ws)
cd ..
source install/setup.bash
ros2 run turtlebot3_rl det_obs
Terminal 5: PyGame Operator Visualizer (Can be used to assign robot to targets by mouse click, checkout in Videos) (Make sure you point to src)
cd ..
source install/setup.bash
cd src
ros2 run turtlebot3_rl global_database_node
Terminal 6: ROS2-RL Environment Bash (the workspace has to be sourced), the terminal will be empty after run, dont worry unless any issues
source install/setup.bash
cd src
./run_environment.sh
Terminal 7: ROS2-Gazebo Environment Bash (the workspace has to be sourced)
source install/setup.bash
cd src
./run_gazebo_testing.sh
Make sure all 4 robots have been assigned coordinates. Also have a look at Terminal 6 if all the 4 robot coordinates have been recieved. This is crucial, some cases some robot coordinates would have not recieved. This can solved by just rerunning the Terminal 6 and 7 in order. This issue is due to ROS2 service-client delay.
Terminal 8: ROS2-Agent Bash (the workspace has to be sourced)
source install/setup.bash
cd src
./run_test_agent.sh
The run_test_agent will load the pretrained weights from models folder. Once that succeeds, one can visulaize the robot swarm move in the Gazebo Engine.
Important: As explained earlier, the user is supposed to uncomment a few lines in few scripts to make the code work for the larger environment
env.py file in turtlebot3_rl package, line 18
gazebo.py file in turtlebot3_rl package, line 21
agent.py file in turtlebot3_rl package, line 27
td3.py file in turtlebot3_rl package, line 13
util.py file in turtlebot3_rl package, line 10
Terminal 1: Spawn 5 bots launch file (Make sure you point to swarm_ws)
cd ..
source install/setup.bash
ros2 launch sim_gazebo spawn_5_bots.launch.py
Terminal 2: ROS2-RL Environment Bash (the workspace has to be sourced), the terminal will be empty after run, dont worry unless any issues
source install/setup.bash
cd src
./run_environment_large.sh
Terminal 3: ROS2-Gazebo Environment Bash (the workspace has to be sourced)
source install/setup.bash
cd src
./run_gazebo_testing_large.sh
Make sure all 5 robots have been assigned coordinates. Also have a look at Terminal 2 if all the 4 robot coordinates have been recieved. This is crucial, some cases some robot coordinates would have not recieved. This can solved by just rerunning the Terminal 2 and 3 in order. This issue is due to ROS2 service-client delay.
Terminal 4: ROS2-Agent Bash (the workspace has to be sourced)
source install/setup.bash
cd src
./run_test_agent_large.sh
The run_test_agent will load the pretrained weights from models folder. Once that succeeds, one can visulaize the robot swarm move in the Gazebo Engine.
- Train the Swarm with other RL algorithms like PPO, MAPPO, MADDPG.
- Turn it into an Decentralized Training and Centralized Execution Problem.
- Code up a Graph-RL scenario for training and execution. <<<<<<< HEAD =======
485312a (Added Videos and GIFs)