simple_controller_pkg

Introduction

This package provides a learning-based cooperative controller for a group of autonomous underwater vehicles (AUVs) operating together in simulation. The controller is designed for cooperative mission execution under uncertain dynamics and includes a neural-network-based learning component to improve coordination, tracking, and adaptability.

The package also includes supporting launch files, configuration files, CUDA-based RBF utilities, and a yaw extraction node used in the control pipeline.

Tested Environment

• ROS 2: Jazzy
• Ubuntu: 24.04

Directory Information

• config
  • swarm_control.yaml: controller and mission-related parameters.
• include
  • rbf_cuda.hpp: header for CUDA-based RBF neural network utilities.
• launch
  • swarm_control.launch.py: launch file for the controller setup.
• src
  • rbf_cuda.cu: CUDA implementation of the RBF neural-network-related computations.
  • simple_controller.cpp: main controller implementation.
  • yaw_extractor_node.cpp: node for extracting desired yaw information from odometry/waypoint topics.
• test
  • sendwaypoints_srv_call.txt: example service call for sending waypoints.

Main Features

• Learning-based cooperative control for multiple agents
• Neural-network-based adaptation for uncertain system dynamics
• Support for single-vehicle and multi-vehicle launch setups
• CUDA-based RBF utility implementation
• Yaw extraction utility for waypoint-following/control integration

Dependencies

This package is intended to be used inside a ROS 2 workspace together with the simulation and MVP-related packages used in the project.

Typical dependencies may include:

• rclcpp
• std_msgs
• geometry_msgs
• nav_msgs
• tf2
• tf2_ros
• tf2_geometry_msgs
• Eigen3
• CUDA (for rbf_cuda.cu)

Please make sure the required ROS 2 packages and system libraries are installed before building.

Building the Workspace

From the root of your ROS 2 workspace:

cd ~/ros2_ws
colcon build --packages-select simple_controller_pkg
source install/setup.bash

Learning Phase Control

Each controller node exposes a runtime parameter called learning_phase. The supported values are:

• learning
• steady_recording
• frozen

The commands below switch all three agents together. They assume the controller nodes are running as:

• /mauv_1/simple_controller_node
• /mauv_2/simple_controller_node
• /mauv_3/simple_controller_node

Start Steady Recording For All Agents

This keeps learning active and starts collecting steady-phase weight samples inside each controller.

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param set /${agent}/simple_controller_node learning_phase steady_recording
done

Freeze Learning For All Agents

This computes the averaged steady-phase weight w_bar for each agent, uploads it to the GPU, and disables further weight updates.

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param set /${agent}/simple_controller_node learning_phase frozen
done

Switch All Agents To Shared Swarm Knowledge

After all three agents have frozen their local w_bar, the swarm_weight_manager_node computes:

Wbar_shared = (Wbar_1 + Wbar_2 + Wbar_3) / 3

Use the command below to tell all agents to replace their local frozen knowledge with the shared swarm-average weight:

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param set /${agent}/simple_controller_node knowledge_source swarm_average
done

Reuse The Saved Shared Weight On A Fresh Run

To start directly with the saved shared swarm weight and no new learning:

ros2 launch simple_controller_pkg swarm_control.launch.py use_saved_shared_weights:=true

To start directly with the saved shared weight in the rotated formation:

ros2 launch simple_controller_pkg swarm_control.launch.py use_saved_shared_weights:=true formation_profile:=rotated

The saved files are:

• ~/.ros/simple_controller/shared_wbar.bin
• ~/.ros/simple_controller/shared_wbar.meta

Return All Agents To Local Frozen Knowledge

If you want to switch back from the shared swarm average to each agent's own local frozen w_bar:

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param set /${agent}/simple_controller_node knowledge_source local_average
done

Rotate The Formation For Transfer Testing

This keeps the mission running, but changes the role/slot assignment so the agents test the learned knowledge in a new configuration:

• mauv_1 takes the old mauv_2 slot
• mauv_2 takes the old mauv_3 slot
• mauv_3 takes the old mauv_1 slot

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param set /${agent}/simple_controller_node formation_profile rotated
done

Return To The Training Formation

To restore the original learned/training configuration:

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param set /${agent}/simple_controller_node formation_profile training
done

Mission Selection

Each controller node also exposes a runtime parameter called mission_profile. The supported values are:

• figure_eight
• quarter_square_hold

The default mission remains figure_eight. The new quarter_square_hold mission moves along a smooth square-corner path and then holds the final point.

Start The Quarter-Square Mission At Launch

ros2 launch simple_controller_pkg swarm_control.launch.py mission_profile:=quarter_square_hold

Switch All Agents To The Quarter-Square Mission At Runtime

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param set /${agent}/simple_controller_node mission_profile quarter_square_hold
done

Return All Agents To The Figure-Eight Mission

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param set /${agent}/simple_controller_node mission_profile figure_eight
done

The quarter-square geometry and timing are configured in config/swarm_control.yaml using:

• quarter_square_origin_x
• quarter_square_origin_y
• quarter_square_side_length
• quarter_square_turn_radius
• quarter_square_leg_time
• quarter_square_turn_time

Return All Agents To Learning Mode

This re-enables online weight updates for all agents.

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param set /${agent}/simple_controller_node learning_phase learning
done

Optional Check

To confirm the current learning phase for all three agents:

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param get /${agent}/simple_controller_node learning_phase
done

To confirm the active knowledge source for all three agents:

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param get /${agent}/simple_controller_node knowledge_source
done

To fully confirm the runtime mode, check both:

for agent in mauv_1 mauv_2 mauv_3; do
  echo -n "${agent} phase: "
  ros2 param get /${agent}/simple_controller_node learning_phase
  echo -n "${agent} source: "
  ros2 param get /${agent}/simple_controller_node knowledge_source
done

To confirm the active formation profile for all three agents:

for agent in mauv_1 mauv_2 mauv_3; do
  ros2 param get /${agent}/simple_controller_node formation_profile
done

Important Note

Do not switch to frozen before running steady_recording long enough to collect samples. If no steady-phase samples were recorded, the node will reject the freeze request and stay in its current mode.