ASPECT:Analogical Semantic Policy Execution via Language Conditioned Transfer

TL;DR

This paper introduces ASPECT, a method that leverages language-conditioned transfer and large language models to enable reinforcement learning agents to generalize to new, complex tasks without retraining.

Contribution

It replaces discrete class systems with language conditioning and uses an LLM as a semantic operator for zero-shot transfer in RL tasks.

Findings

Achieves zero-shot transfer across diverse complex tasks.

Enables policy reuse through language-conditioned state representations.

Outperforms previous methods constrained by fixed categories.

Abstract

Reinforcement Learning (RL) agents often struggle to generalize knowledge to new tasks, even those structurally similar to ones they have mastered. Although recent approaches have attempted to mitigate this issue via zero-shot transfer, they are often constrained by predefined, discrete class systems, limiting their adaptability to novel or compositional task variations. We propose a significantly more generalized approach, replacing discrete latent variables with natural language conditioning via a text-conditioned Variational Autoencoder (VAE). Our core innovation utilizes a Large Language Model (LLM) as a dynamic \textit{semantic operator} at test time. Rather than relying on rigid rules, our agent queries the LLM to semantically remap the description of the current observation to align with the source task. This source-aligned caption conditions the VAE to generate an imagined state compatible with the agent's original training, enabling direct policy reuse. By harnessing the flexible reasoning capabilities of LLMs, our approach achieves zero-shot transfer across a broad spectrum of complex and truly novel analogous tasks, moving beyond the limitations of fixed category mappings. Code and videos are available \href{https://anonymous.4open.science/r/ASPECT-85C3/}{here}.