Mapping representations in Reinforcement Learning via Semantic Alignment for Zero-Shot Stitching

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Ricciardi, Antonio Pio, Maiorca, Valentino, Moschella, Luca, Marin, Riccardo, Rodolà, Emanuele
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866915181195427840
author Ricciardi, Antonio Pio
Maiorca, Valentino
Moschella, Luca
Marin, Riccardo
Rodolà, Emanuele
author_facet Ricciardi, Antonio Pio
Maiorca, Valentino
Moschella, Luca
Marin, Riccardo
Rodolà, Emanuele
contents Deep Reinforcement Learning (RL) models often fail to generalize when even small changes occur in the environment's observations or task requirements. Addressing these shifts typically requires costly retraining, limiting the reusability of learned policies. In this paper, we build on recent work in semantic alignment to propose a zero-shot method for mapping between latent spaces across different agents trained on different visual and task variations. Specifically, we learn a transformation that maps embeddings from one agent's encoder to another agent's encoder without further fine-tuning. Our approach relies on a small set of "anchor" observations that are semantically aligned, which we use to estimate an affine or orthogonal transform. Once the transformation is found, an existing controller trained for one domain can interpret embeddings from a different (existing) encoder in a zero-shot fashion, skipping additional trainings. We empirically demonstrate that our framework preserves high performance under visual and task domain shifts. We empirically demonstrate zero-shot stitching performance on the CarRacing environment with changing background and task. By allowing modular re-assembly of existing policies, it paves the way for more robust, compositional RL in dynamically changing environments.
format Preprint
id arxiv_https___arxiv_org_abs_2503_01881
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Mapping representations in Reinforcement Learning via Semantic Alignment for Zero-Shot Stitching
Ricciardi, Antonio Pio
Maiorca, Valentino
Moschella, Luca
Marin, Riccardo
Rodolà, Emanuele
Machine Learning
Artificial Intelligence
68T07
I.2.6
Deep Reinforcement Learning (RL) models often fail to generalize when even small changes occur in the environment's observations or task requirements. Addressing these shifts typically requires costly retraining, limiting the reusability of learned policies. In this paper, we build on recent work in semantic alignment to propose a zero-shot method for mapping between latent spaces across different agents trained on different visual and task variations. Specifically, we learn a transformation that maps embeddings from one agent's encoder to another agent's encoder without further fine-tuning. Our approach relies on a small set of "anchor" observations that are semantically aligned, which we use to estimate an affine or orthogonal transform. Once the transformation is found, an existing controller trained for one domain can interpret embeddings from a different (existing) encoder in a zero-shot fashion, skipping additional trainings. We empirically demonstrate that our framework preserves high performance under visual and task domain shifts. We empirically demonstrate zero-shot stitching performance on the CarRacing environment with changing background and task. By allowing modular re-assembly of existing policies, it paves the way for more robust, compositional RL in dynamically changing environments.
title Mapping representations in Reinforcement Learning via Semantic Alignment for Zero-Shot Stitching
topic Machine Learning
Artificial Intelligence
68T07
I.2.6
url https://arxiv.org/abs/2503.01881