Find Object skill for AgenticROS: the robot rotates in place until it sees a target object in the camera feed, then stops.

• Behavior: One-shot search. The skill rotates the robot in place (clockwise by default) and runs YOLOv8n on each captured camera frame. As soon as it detects an instance of the requested COCO class above a confidence threshold, it stops the robot and returns the bounding box plus the horizontal offset of the object from image center.
• Targets: Any of the 80 COCO classes — e.g. bottle, cup, vase, chair, cell phone, laptop, bowl, book, tv, couch. Common aliases (phone, tv/television, sofa, plant, bike) are accepted.
• No ROS2 node required: Detection runs entirely inside the OpenClaw gateway process. The skill subscribes to the camera topic via the AgenticROS transport and publishes geometry_msgs/Twist to cmd_vel.
• Stops on success or timeout: If the timeout elapses without detection, the skill stops the robot and returns found: false with the elapsed time.
• Tools:
  • find_object — Rotate and search for a target object. Returns whether it was found, confidence, bounding box, and horizontal offset (-1 left edge, 0 center, +1 right edge).
• Config: All options live under config.skills.find (see Config options).

1. Install the skill where the OpenClaw gateway can load it:
   • Option A (recommended): Add to AgenticROS config as a package name:
     • In OpenClaw config, under plugins.entries.agenticros.config, set:
       • skillPackages: ["agenticros-skill-find"]
     • Install the package in the same environment as the gateway (e.g. pnpm add agenticros-skill-find in the gateway app).
   • Option B: Clone this repo into a directory and add that directory to skillPaths:
     • skillPaths: ["/path/to/agenticros-skill-find"]
     • Run pnpm install and pnpm build in this repo.
2. Configure the skill in the same config under skills.find (see Config options).
3. Restart the OpenClaw gateway so the plugin loads the skill.
4. In chat, use natural language: e.g. "find the bottle", "look for a chair", "locate my phone". The agent will call find_object with the appropriate target.

On its first call, the skill downloads yolov8n.onnx (~6 MB) from the public Ultralytics mirror to ~/.agenticros/models/yolov8n.onnx. Subsequent calls reuse the cached file.

To override:

• AGENTICROS_YOLOV8_MODEL — absolute path to a local .onnx file.
• AGENTICROS_YOLOV8_URL — alternative download URL.

The skill gets transport from the plugin via context:

• context.getTransport() — ROS2 transport (subscribe/publish).
• context.logger — Plugin logger.

Config is the full AgenticROS config; the skill only uses config.skills.find, plus config.robot.namespace / config.teleop.cmdVelTopic / config.safety.maxAngularVelocity for topic resolution and safety clamping.

Returns an object with found, target, classId, elapsedSeconds, rotationDirection, angularSpeed, and (when found) a detection block: confidence, cx, cy, width, height, imageWidth, imageHeight, horizontalOffset.

• Package: package.json has "agenticrosSkill": true and a main entry that exports registerSkill(api, config, context).
• Config: Skill-specific options live under config.skills.find. The skill validates and defaults its own slice.
• Context: Use context.getTransport() for ROS2 and context.logger for logging.
• Registration: Call api.registerTool(...) inside registerSkill. Depend only on the public skill API and types exported by @agenticros/core.

For the full contract and how to create a third-party skill, see docs/skills.md in the AgenticROS repo, and use agenticros-skill-followme as the reference template.

YOLOv8n model from Ultralytics.