CODE-SHARP: AI Framework for Open-Ended Skill Discovery
Researchers have introduced CODE-SHARP (Continuous Open-ended Discovery and Evolution of Skills as Hierarchical Reward Programs), a framework that uses Foundation Models to autonomously grow and evolve a library of Python programs encoding skills. These programs, called SHARPs, define local success conditions and prerequisites linked to previously discovered skills. At runtime, SHARPs dynamically route an agent through tasks via reinforcement learning, starting from scratch and using only source code. The approach aims to reduce human-in-the-loop engineering, enabling transferability to novel environments. The paper is available on arXiv under identifier 2602.10085.
Key facts
- CODE-SHARP stands for Continuous Open-ended Discovery and Evolution of Skills as Hierarchical Reward Programs.
- The framework leverages Foundation Models (FMs).
- Skills are encoded as Python programs called SHARPs.
- Each SHARP contains a local success condition and prerequisites.
- SHARPs route agents dynamically through tasks.
- The agent is trained via reinforcement learning from scratch.
- The approach minimizes human-in-the-loop engineering.
- The paper is published on arXiv (identifier 2602.10085).
Entities
Institutions
- arXiv