Language Guided Skill Discovery
Seungeun Rho, Laura Smith, Tianyu Li, Sergey Levine, Xue Bin Peng,, Sehoon Ha
TL;DR
This paper introduces Language Guided Skill Discovery (LGSD), a framework that uses language prompts and large language models to generate semantically diverse skills for agents, improving downstream task utility and interpretability.
Contribution
LGSD leverages LLMs to directly maximize semantic diversity of skills, enabling targeted skill discovery guided by natural language prompts.
Findings
LGSD enables robots to visit different areas based on prompts.
Language guidance improves skill diversity over existing methods.
LGSD allows natural language utilization of learned skills.
Abstract
Skill discovery methods enable agents to learn diverse emergent behaviors without explicit rewards. To make learned skills useful for unknown downstream tasks, obtaining a semantically diverse repertoire of skills is essential. While some approaches introduce a discriminator to distinguish skills and others aim to increase state coverage, no existing work directly addresses the "semantic diversity" of skills. We hypothesize that leveraging the semantic knowledge of large language models (LLMs) can lead us to improve semantic diversity of resulting behaviors. In this sense, we introduce Language Guided Skill Discovery (LGSD), a skill discovery framework that aims to directly maximize the semantic diversity between skills. LGSD takes user prompts as input and outputs a set of semantically distinctive skills. The prompts serve as a means to constrain the search space into a semantically…
Peer Reviews
Decision·ICLR 2025 Poster
- A theoretical proof of language-distance as a valid pseudometric is provided. - Both locomotion and manipulation tasks have been demonstrated, showing the effectiveness of the proposed approach. - The paper is well presented.
- More complex scenarios should be designed to make the use of LLMs necessary rather than simple positions. - The lack of real-world experiments makes the proposed approach less convincing. - A video showing the robot in action should be included to give readers an intuitive understanding of learning performance.
- The method tried to link skill states with their semantic meaning, resulting in meaningful and language-controllable skills. - LGSD provides a theoretical foundation for the language-distance metric as a pseudometric, detailing how it can approximate semantic diversity.
- The experimental scenarios are simple, in which the exampled prompts and semantically controlled spaces are easy to follow yet fail to demonstrate the generalizablity and scalability --- after all, the method relies much on the description of states. LGSD’s dependence on LLMs for real-time distance evaluation might limit scalability to complex, real-time environments. - As I understand, users have to provide specified "skill constraints" (via prompts, such as "move north" etc.), then how can
- Interesting and novel contribution - while using LLMs is not rare, this is an interesting way of using them for encouraging skill discovery. - The proposed method is well-supported mathematically. - Good zero-shot language-specified goal tracking with the learned skill predictor network.
- Not fully convinced that LLMs are necessary for the proposed scenarios. Showing the performance on more complex situations where the semantic understanding of LLMs is necessary would make the case much stronger. - More qualitative evaluation (i.e. videos) would be good for evaluating the learned skills.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsNatural Language Processing Techniques · Topic Modeling · Educational Assessment and Pedagogy
MethodsSparse Evolutionary Training