vision_pipeline¶

Role: Meta-package that orchestrates the perception stack. Bundles launch files and provides lifecycle + composable-node managers so a single service call can flip a whole task pipeline on or off.

Key entrypoints¶

File	Purpose
`launch/gate.launch.py`, `bin.launch.py`, `torpedo.launch.py`, `slalom.launch.py`, `trash.launch.py`, `image_matching.launch.py`	One launch per RoboSub task.
`vision_pipeline/lifecycle_manager_node.py`	Drives `LifecycleNode` state transitions (`configure` → `activate`).
`vision_pipeline/component_manager_node.py`	Loads / unloads composable nodes inside a container.
`config/auv4_orin/*.yaml`	Per-task model / topic configs.

ROS interfaces¶

Type	Name
Service	`manage_nodes` (`lifecycle_msgs/srv/ChangeState`)
Service	`manage_components` (`std_srvs/srv/SetBool`)

Vision nodes are namespaced /auv4/<task>/... (preserved verbatim from upstream).

Pipeline composition¶

This package itself owns no perception logic. It composes:

yolo_ros_trt: segmentation
pose_estimator. PnP → TF
image_matching. XFeat correspondences
image_processing: brighten / restamp
filters: clustering / SORT3D
aruco_loco, depth_anything_ros2_trt, custom_image_republisher. Auxiliary

Gotchas¶

Lifecycle state

Lifecycle nodes start inactive. They must be activated explicitly via manage_nodes or ros2 lifecycle set /<node> activate before they publish anything.

Depth Anything is .onnx, not .engine

Depth Anything configs must reference .onnx files. The runtime compiles .engine on first run.

QoS mismatch

Image topics use sensor_data QoS (best-effort). Subscribers expecting reliable QoS will not receive messages. This is the most common silent-perception-failure mode.

All vision nodes run on the Orin in the real robot, consuming decompressed images that the SBC republishes (orin_cam_repub.launch.py).