Steve-Evolving: Open-World Embodied Self-Evolution via Fine-Grained Diagnosis and Dual-Track Knowledge Distillation

Researchers introduce Steve-Evolving, a new AI framework for open-world embodied agents that uses fine-grained diagnosis and knowledge distillation to improve long-horizon task performance. The system organizes interaction experiences into structured tuples and continuously evolves without model parameter updates, showing improvements in Minecraft testing environments.

• →Steve-Evolving is a non-parametric self-evolving framework that couples execution diagnosis with dual-track knowledge distillation for embodied AI agents.
• →The system uses a three-phase approach: Experience Anchoring, Experience Distillation, and Knowledge-Driven Closed-Loop Control.
• →Experiences are organized in a three-tier space with multi-dimensional indices including condition signatures, spatial hashing, and semantic tags.
• →Successful trajectories are generalized into reusable skills while failures become executable guardrails to prevent risky operations.
• →Testing on Minecraft MCU long-horizon tasks demonstrated consistent improvements over static-retrieval baseline methods.