Teaching General-Purpose Agents to Act on Earth via Skill Evolution
A Self-evolving Skill Framework for Tool-Driven Geospatial Intelligence
Haojun Tang1,2 Β· Wenda Zhao2,β Β· Xiang Zhou2 Β· Qiwei Lin3 Β· Lida Huang1 Β· Yan Wang1,β
1 Institute for AI Industry Research (AIR), Tsinghua University
2 Dalian University of Technology
3 Beijing Institute of Radio Measurement
β Corresponding authors
Figure 1. Teaser of SkillEarth. (a) Skill-guided execution examples showing how extracted skills guide multi-step tool execution. (b) Quantitative comparisons with existing RS agents on OpenEarth-Bench (End-to-End mode) and Earth-Bench (IF mode).
SkillEarth SkillBank for Earth-Bench
This repository releases the Earth-Bench SkillBank from SkillEarth: Teaching General-Purpose Agents to Act on Earth via Skill Evolution.
The bank contains reusable SKILL.md files for Earth-Bench tool-use workflows, including spectral, product, and RGB tasks. Each skill records when it applies, an execution plan, grounded observation routing rules, and the required tool interfaces.
Included SkillBank
| Directory | Backbone | Benchmark | Supported modes |
|---|---|---|---|
qwen3-4b-instruct-2507/ |
Qwen3-4B-Instruct | Earth-Bench | Automatic Planning (AP) and Instruction Following (IF) |
Directory Layout
qwen3-4b-instruct-2507/
βββ <skill-name>/
β βββ SKILL.md
βββ ...
Use with SkillEarth
Download this repository, then place the model directory where the Earth-Bench runner can find it:
cd /path/to/SkillEarth
cp -a /path/to/downloaded-skillbank-eb/qwen3-4b-instruct-2507 \
skillbank/qwen3-4b-instruct-2507
Citation
If you use this SkillBank, please cite the SkillEarth paper.