A multimodal grounding skill that utilizes GLM-4.7V to detect and locate objects, text, or UI elements in images with precise bounding boxes.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install grounding-anything
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install grounding-anything using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Grounding Multimodal Object Detection?
The Grounding skill empowers Openclaw Skills with advanced spatial reasoning by leveraging the GLM-4.7V multimodal model. It enables an agent to analyze image data and identify the exact coordinates of requested objects, providing normalized bounding boxes that can be used for further automation or visual confirmation.
This skill is particularly valuable for developers building vision-enabled agents that need to provide more than just descriptive text. By returning coordinate data, it bridges the gap between high-level visual understanding and low-level pixel manipulation, making it a vital component for any complex vision workflow within Openclaw Skills.
Grounding Multimodal Object Detection Use Cases
- Automatically detecting and boxing UI elements in web or mobile screenshots for testing.
- Locating specific objects like products, people, or items in a photograph upon user request.
- Extracting spatial data for OCR tasks to identify where text is located within a document.
- Visualizing AI reasoning by drawing bounding boxes around identified regions of interest.
How Grounding Multimodal Object Detection Works
- The skill receives a user input consisting of an image path and a prompt requesting specific object detection.
- It executes a call to the model API using the HttpInterface, passing the image and a specialized prompt for coordinate output.
- The model returns a text response containing normalized coordinates in the [xmin, ymin, xmax, ymax] format.
- A specialized parser extracts these coordinates from various bracket styles and flattens nested lists.
- The skill converts the normalized 0-1000 scale back into pixel coordinates based on the original image dimensions.
- Finally, it uses a visualization utility to draw the bounding boxes and labels onto the image, saving the result for the user.
Grounding Multimodal Object Detection Setup
To integrate this capability into your Openclaw Skills environment, you must first configure your model host. Ensure your model API address is correctly set in /root/.openclaw/agents/main/agent/models.json.
Because model requests often happen over internal networks, you must set your proxy environment variables:
export NO_PROXY='your-model-host-ip'
export no_proxy='your-model-host-ip'
You can then call the skill using the HttpInterface and utility functions provided in the package.
Grounding Multimodal Object Detection Data Schema & Taxonomy
The Grounding skill processes and generates data according to the following structure:
| Component | Type | Format/Value |
|---|---|---|
| Input Image | File Path | Local path to the source image |
| Normalized Coords | Array | [xmin, ymin, xmax, ymax] (0-1000 integers) |
| Pixel Coords | Array | Calculated as pixel = coord * img_dimension / 1000 |
| Visual Output | Image File | Saved JPEG/PNG with rendered boxes and labels |
| API Config | JSON | Managed in the agent's models.json configuration |
Grounding Multimodal Object Detection Advanced Features
- Robust coordinate parsing that supports multiple bracket styles including standard brackets, curly braces, and custom
tags. - Automatic context window management that checks for truncated responses to ensure no bounding boxes are missed.
- Toggleable reasoning mode using the no_think parameter to prioritize detection speed over chain-of-thought processing.
- Highly customizable visualization engine allowing for specific color-coding, label assignment, and line thickness per object.
SKILL.md
Loading
Related Openclaw Skills
LinkedIn AI Roast Generator
An automated skill that scrapes LinkedIn profiles to generate humorous, savage audio roasts regarding professional AI replacement.
Traffic Monitor Skill
A specialized network monitoring skill designed to track monthly bandwidth consumption and provide automated traffic alerts.
FTShare-cb-data
A comprehensive data skill set for querying Chinese convertible bond lists, fundamental information, and historical market trends.
QMD - Quick Markdown Search
A high-performance local search engine that brings advanced RAG capabilities to your markdown knowledge bases and notes.
Package Version Tracker
A streamlined utility for developers to query, track, and compare package versions from npm and PyPI registries.
China Hot Trends Aggregator
A comprehensive data aggregator that fetches real-time hot searches and trending topics from China's leading social, tech, and financial platforms.
Featured*
