caatingarobotics

Fork of caatingarobotics retargeted to ROS 2 Jazzy + Gazebo Harmonic, for the Sowbot platform.

The upstream project targets ROS 2 Humble / earlier Gazebo. This fork ports the workspace to Jazzy (Ubuntu 24.04 Noble) and Gazebo Harmonic, and adds the sowbot_row_follow package for monocular crop-row following.

⚠️ Experimental. The Jazzy/Harmonic port is a work in progress and has not been validated on physical hardware.

How this repo is used — read this first

You do not clone or build this repository directly. It is consumed by the feldfreund_devkit_ros devkit, which is the supported way to build and run everything.

The devkit's docker/Dockerfile clones this fork automatically during the image build:

git clone --depth 1 -b main \
  https://github.com/samuk/caatingarobotics src/caatingarobotics

All four packages here are then built by colcon inside the devkit's ros:jazzy-based Docker image, alongside the rest of the Sowbot stack (devkit_launch, devkit_driver, fusioncore, topological_navigation, virtual_maize_field, etc.).

Practical consequence: to pick up changes made in this fork, push them to the main branch and rebuild the devkit image.

Running (via the devkit)

Everything runs inside the devkit's Docker container, orchestrated by its manage.py. In short:

# 1. Clone the devkit (not this repo)
git clone https://github.com/Agroecology-Lab/feldfreund_devkit_ros.git
cd feldfreund_devkit_ros

# 2. Build the image — the Dockerfile pulls this fork in automatically.
#    Add +sim to also install Gazebo Harmonic (needed for the simulation).
./manage.py build           # or:  ./manage.py build +sim
#    ./manage.py full-build  # --no-cache rebuild

# 3. Run
./manage.py                 # sim / limbic stack — exercises agro_robot_sim
./manage.py neo             # Neo row-follow stack — runs sowbot_row_follow

./manage.py neo launches the devkit's neo.launch.py, which delegates to sowbot_row_follow/launch/crop_row_nav.launch.py — i.e. the crop_row_node from this repo. Launch arguments pass straight through, e.g.:

./manage.py neo camera_index:=2

For full setup (the .env hardware-detection step, USB/GNSS handling, the NiceGUI cockpit, sim vs. hardware mode, etc.) see the devkit README. This README only covers what the caatingarobotics packages contribute.

Packages in this repo

The devkit image builds all four packages, but the current Sowbot work only launches two of them:

Package	Purpose	Status
agro_robot_sim	Robot URDF/Xacro, Gazebo Harmonic worlds, simulation assets	Active — used by the sim stack
sowbot_row_follow	Monocular RGB crop-row following (crop_row_node, limbic_row_follow)	Active — used by ./manage.py neo
caatinga_vision	Vision / inference pipeline	Not used yet — candidate for future work
caatinga_nav	Standalone Nav2 navigation stack	Not used (the devkit uses LCAS topological_navigation instead)

sowbot_row_follow

Monocular RGB crop-row following, derived from visual-multi-crop-row-navigation (Agricultural-Robotics-Bonn, BSD-2-Clause). ExG vegetation index + Otsu threshold, scan-window line fitting, Cherubini & Chaumette visual servoing. Publishes /aoc/conditions/row_offset, /aoc/conditions/row_heading_error, and the /aoc/heartbeat/neo_vision perception heartbeat into the AOC navigation layer.

It contains two nodes:

• crop_row_node — runs on the Neo SBC; opens the camera directly and does the row-following. This is what ./manage.py neo runs.
• limbic_row_follow — runs on Limbic as a Nav2 NavigateToPose action server, enabling/monitoring Neo over the crossover link.

Camera calibration is mandatory before any field use — set camera_height_m and camera_tilt_deg in src/sowbot_row_follow/config/crop_row_params.yaml. See that package's README.md for the full parameter and tuning notes.

agro_robot_sim

Robot description (Xacro/URDF), Gazebo Harmonic worlds, and simulation assets for the Caatinga agro-robot. In the devkit's sim mode the robot model from this package is spawned into a generated virtual_maize_field world. Requires the devkit image to be built with +sim.

Repository structure

caatingarobotics/
├── datasets/
│   └── agro_v1/data.yaml
├── src/
│   ├── agro_robot_sim/        # Active — simulation (URDF, worlds, launch)
│   │   ├── launch/
│   │   ├── config/
│   │   ├── maps/
│   │   ├── urdf/
│   │   └── worlds/
│   ├── sowbot_row_follow/     # Active — monocular crop-row following
│   │   ├── sowbot_row_follow/
│   │   ├── launch/
│   │   └── config/
│   ├── caatinga_vision/       # Not used yet — vision/inference pipeline
│   └── caatinga_nav/          # Not used — standalone Nav2 stack
├── CHANGELOG.md
└── README.md

Developing a package standalone

The supported run path is the devkit image above. If you need to iterate on a single package in isolation, you can still build it in a local ROS 2 Jazzy environment:

source /opt/ros/jazzy/setup.bash
rosdep install --from-paths src/sowbot_row_follow --ignore-src -r -y
colcon build --symlink-install --packages-select sowbot_row_follow

Note this does not reproduce the devkit runtime (CycloneDDS config, drivers, launch orchestration). Rebuild the devkit image to run changes for real.

Licensing

Licensing is mixed across the workspace:

• agro_robot_sim, caatinga_vision, caatinga_nav — Apache-2.0
• sowbot_row_follow — BSD-2-Clause (derived from the Bonn package; see its README for full attribution)

Notes and future work

• The CI workflow under .github/workflows/ still targets ROS 2 Humble and has not been updated for the Jazzy port.
• The devkit image already ships a CPU PyTorch + YOLOX runtime and YOLOX-Nano weights, so revisiting caatinga_vision for perception work later should not need new dependencies.
• cv_bridge must match the system NumPy ABI. A runtime error like "a module compiled using NumPy 1.x cannot be run in NumPy 2.x" means the cv_bridge build and the runtime numpy disagree — this is resolved in the devkit's Docker image, not in this repo.