MNTFields TurtleBot bundle

This bundle contains the MNTFields ROS 2 Python package and the checked-in TurtleBot model data needed for the included navigation model. It is intended for a TurtleBot3 simulation first; the Jackal notes at the end describe the changes required before using the same stack on a Jackal.

Contents

• src/mntfields: MNTFields mapping, learning, navigation, sensing, launch, configuration, and RViz code.
• models/mntfields_model: region checkpoints, map, topology, and room graph used by the navigation launch.

The SDF local replanner is included in the package, but its external comp-sdf-dyn-nav repository and SDF checkpoints are not copied here. Run the standard MNTFields navigation first; enable the SDF replanner only after that repository and its checkpoints are installed separately.

TurtleBot3 simulation

Use a ROS 2 Humble environment with the TurtleBot3, Gazebo, Nav2, SLAM Toolbox, and ROS 2 Python dependencies installed. From the extracted bundle directory:

source /opt/ros/humble/setup.bash
export TURTLEBOT3_MODEL=waffle_pi
colcon build --symlink-install --packages-select mntfields
source install/setup.bash
ros2 launch mntfields mntfields_mapping.launch.py

The mapping launch starts Gazebo, the TurtleBot3, SLAM Toolbox, and the MNTFields grid-map node. For the included saved model and map, use a second terminal:

source /opt/ros/humble/setup.bash
source install/setup.bash
ros2 launch mntfields mntfields_navigation.launch.py \
  model_dir:=$PWD/models/mntfields_model \
  map_yaml:=$PWD/models/mntfields_model/map.yaml

The launch file currently starts its own simulation processes, so normally run either mapping or navigation, not both at once. Set TURTLEBOT3_MODEL=burger if using the Burger model instead of the default Waffle Pi.

Single-field mode without room segmentation

MNTFields can run as one continuously trained field instead of splitting the environment into room regions. Use antfields_gridmap_node, not the room-segmenting mapping_node or gridmap_node paths:

source /opt/ros/humble/setup.bash
source install/setup.bash
export TURTLEBOT3_MODEL=waffle_pi
ros2 launch mntfields antfields_sim_mapping.launch.py

The single-region implementation is:

• src/mntfields/mntfields/mntfields_ros/mapping/antfields_gridmap_node.py
• src/mntfields/launch/antfields_sim_mapping.launch.py
• src/mntfields/launch/antfields_tb4_mapping.launch.py

These paths do not use room segmentation, room ownership, or room-based region switching. They create one Region and train one field network. Avoid mntfields_mapping.launch.py, line_mapping.launch.py, mapping_node, and gridmap_node when a single field is desired.

For a larger operating area, edit antfields_gridmap_node.py and increase the single-field bounds:

# Jackal/single-field deployment: cover the complete operating environment.
self.declare_parameter("fixed_region_size_m", 50.0)
self.declare_parameter("expand_region_with_map", True)

The default is a 20 m fixed region with expansion disabled. A larger field requires more GPU memory and may train more slowly as the explored area grows. For the Jackal, also replace the TurtleBot3 driver/description launch, verify /scan, /cmd_vel, base_link/base_footprint, and localization TF, then retune lidar height, velocity limits, and footprint-related parameters.

Real TurtleBot3

Do not start Gazebo. Bring up the robot's drivers, laser scanner, odometry, localization/SLAM, and TF tree first. The MNTFields node expects a LaserScan on /scan, a usable map -> base_link or map -> base_footprint transform, and publishes velocity through /cmd_vel in its navigation path.

Pass the real map/model paths explicitly and set robot_type:=real on the mapping node or equivalent launch configuration. Start at low velocity and verify the scan frame and base_link TF in RViz before enabling autonomous motion.

Jackal migration notes

This code is not automatically Jackal-ready. The following must be adapted:

1. Replace the TurtleBot3 Gazebo launch, URDF, AMCL parameters, and robot model with the Jackal ROS 2 description/driver launch.
2. Confirm the Jackal topics and frames. At minimum, check /scan, /cmd_vel, base_link, base_footprint, odom, and the global localization frame. Override base_frame, cmd_vel_topic, and lidar_scan_topic where necessary.
3. Re-tune lidar_height_m, velocity limits, region radius, path commit distance, and obstacle-clearance parameters for the larger Jackal chassis.
4. For SDF replanning, provide a Jackal body SDF representation. The optimizer needs a Jackal-specific DeepSDF latent code plus its matching normalization JSON; a Gazebo .sdf file alone is not sufficient. Replace the current TurtleBot latent checkpoint and update the SDF node's hard-coded checkpoint path before enabling use_sdf_planner:=true.
5. Test in Jackal simulation with conservative speeds, then validate the footprint against the physical robot including payload, bumpers, and permanently mounted sensors.

The included mntfields_navigation.launch.py is deliberately TurtleBot3/Gazebo specific. Treat it as a reference when writing a Jackal launch file rather than running it unchanged on the Jackal.