SILG

SILG是一个多环境符号互动语言接地基准，整合了多种不同的接地语言学习环境。该数据集包括需要泛化到新动态、实体和部分观察世界的网格世界环境（如RTFM, Messenger, NetHack），以及需要解释复杂场景中丰富自然语言的视觉世界的符号对应物（如ALFWorld, Touchdown）。这些环境共同提供了观察空间、动作空间、语言规范和计划复杂性方面的多样性接地挑战。SILG旨在为研究者提供一个易于使用的基准，以快速评估跨所有这些环境及其各自挑战的方法。

SILG is a multi-environment symbolic interactive language grounding benchmark that integrates a variety of distinct grounded language learning environments. This dataset includes grid-world environments (e.g., RTFM, Messenger, NetHack) that require generalization to novel dynamics, entities, and partially observed worlds, as well as symbolic counterparts of visual worlds that demand interpreting rich natural language in complex scenarios (e.g., ALFWorld, Touchdown). Collectively, these environments offer diverse grounding challenges across observation spaces, action spaces, language specifications, and planning complexity. SILG aims to provide researchers with an easy-to-use benchmark for rapidly evaluating methods across all these environments and their respective challenges.

搜集汇总

背景与挑战

背景概述

SILG是一个多环境符号互动语言接地基准，整合了网格世界和视觉世界等多种接地语言学习环境，旨在提供多样性接地挑战，帮助研究者快速评估跨环境方法。