Reproduce “Neuro-Symbolic Skill Discovery for Conditional Multi-Level Planning” on OpenResearch

About this paper

Hakan Aktas, Yigit Yildirim, Ahmet Firat Gamsiz, Deniz Bilge Akkoc, Erhan Oztop, Emre Ugur

Aktas et al. present a neuro-symbolic framework that autonomously discovers high-level symbolic skills from limited, unlabeled low-level robot demonstrations via a conditional neural architecture with vector quantization. This system supports multi-level planning, leveraging LLMs for high-level task generation and a gradient-based approach for low-level execution, demonstrating enhanced low-level planning performance through self-supervision and achieving up to 86.10% high-level planning success with Gemini and DINO.

Key insights

• A conditional neural architecture with vector quantization can effectively cluster unlabeled low-level demonstrations into distinct high-level symbolic skills without prior knowledge.
• Retraining the skill discovery model using self-discovered cluster labels dramatically improves the performance and generalization capabilities of low-level planning, even with minimal demonstrations (e.g., 2 per skill).
• Integrating LLMs with a learned neuro-symbolic abstraction layer facilitates robust high-level task planning, with external vision models like DINO enhancing environmental analysis for LLM reasoning.

Summary by alphaXiv.