HARRIER ships a single ROS2 node that consumes EO + IR image topics and
publishes detections + diagnostics. It follows
anima_module.yaml:runtime.backends so the same code runs on the GPU
server (CUDA) and on Apple Silicon (MLX).
| Direction |
Topic |
Type |
Notes |
| input |
/harrier/eo/image_raw |
sensor_msgs/Image |
Electro-optical frames |
| input |
/harrier/ir/image_raw |
sensor_msgs/Image |
Infrared frames |
| output |
/harrier/detections |
vision_msgs/Detection2DArray |
Modality-tagged detections |
| output |
/harrier/diagnostics |
diagnostic_msgs/DiagnosticArray |
Backend + latency report |
| output |
/harrier/metadata |
std_msgs/String |
Model name, version, weights hash |
| Name |
Source |
Default |
Purpose |
weights |
launch arg / HARRIER_WEIGHTS |
required |
Path to the YOLO26 best.pt. |
backend |
launch arg / HARRIER_BACKEND |
auto |
Backend hint forwarded to anima_harrier.device. |
ros2 launch harrier harrier_launch.py \
weights:=/mnt/artifacts-datai/models/project_harrier/latest/best.pt \
backend:=cuda
ros/harrier_node.py defers rclpy imports into main() so the
module is testable on any host. handle_frame() is the pure-Python
hook the ROS subscribers delegate to.
- The modality tag is preserved from the topic all the way to the
output
Detection2DArray so downstream stacks can reason about EO
vs IR behaviour per the paper's §IV.A findings.
- Missing weights are treated as a non-fatal warning; HARRIER drops
the frame and lets the supervisor restart it once the artifact is
available.