Saved in:
Bibliographic Details
Main Authors: Lin, Mingxian, Huang, Wei, Li, Yitang, Jiang, Chengjie, Wu, Kui, Zhong, Fangwei, Qian, Shengju, Wang, Xin, Qi, Xiaojuan
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2507.10548
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909689065766912
author Lin, Mingxian
Huang, Wei
Li, Yitang
Jiang, Chengjie
Wu, Kui
Zhong, Fangwei
Qian, Shengju
Wang, Xin
Qi, Xiaojuan
author_facet Lin, Mingxian
Huang, Wei
Li, Yitang
Jiang, Chengjie
Wu, Kui
Zhong, Fangwei
Qian, Shengju
Wang, Xin
Qi, Xiaojuan
contents Recent advanced vision-language models(VLMs) have demonstrated strong performance on passive, offline image and video understanding tasks. However, their effectiveness in embodied settings, which require online interaction and active scene understanding remains limited. In such scenarios, an agent perceives the environment from a first-person perspective, with each action dynamically shaping subsequent observations. Even state-of-the-art models such as GPT-4o, Claude 3.5 Sonnet, and Gemini 2.5 Pro struggle in open-environment interactions, exhibiting clear limitations in spatial reasoning and long-horizon planning. To address this gap, we introduce EmRACE-3K, a dataset of over 3,000 language-guided tasks situated in diverse, photorealistic environments constructed using Unreal Engine and the UnrealCV-Zoo framework. The tasks encompass a wide range of embodied challenges, including navigation, object manipulation, and multi-stage goal execution. Each task unfolds as a multi-step trajectory, pairing first-person visual observations with high-level instructions, grounded actions, and natural language rationales that express the agent's intent at every step. Using EmRACE-3K, we establish a benchmark to evaluate the embodied reasoning capabilities of VLMs across three key dimensions: Exploration, Dynamic Spatial-Semantic Reasoning, and Multi-stage Goal Execution. In zero-shot settings, all models achieve success rates below 20%, underscoring the challenge posed by our benchmark and the current limitations of VLMs in interactive environments. To demonstrate the utility of EmRACE-3K, we further fine-tune Qwen2.5-VL-7B using supervised learning followed by reinforcement learning. This approach yields substantial improvements across all three challenge categories, highlighting the dataset's effectiveness in enabling the development of embodied reasoning capabilities.
format Preprint
id arxiv_https___arxiv_org_abs_2507_10548
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle EmbRACE-3K: Embodied Reasoning and Action in Complex Environments
Lin, Mingxian
Huang, Wei
Li, Yitang
Jiang, Chengjie
Wu, Kui
Zhong, Fangwei
Qian, Shengju
Wang, Xin
Qi, Xiaojuan
Computer Vision and Pattern Recognition
Artificial Intelligence
Computation and Language
Recent advanced vision-language models(VLMs) have demonstrated strong performance on passive, offline image and video understanding tasks. However, their effectiveness in embodied settings, which require online interaction and active scene understanding remains limited. In such scenarios, an agent perceives the environment from a first-person perspective, with each action dynamically shaping subsequent observations. Even state-of-the-art models such as GPT-4o, Claude 3.5 Sonnet, and Gemini 2.5 Pro struggle in open-environment interactions, exhibiting clear limitations in spatial reasoning and long-horizon planning. To address this gap, we introduce EmRACE-3K, a dataset of over 3,000 language-guided tasks situated in diverse, photorealistic environments constructed using Unreal Engine and the UnrealCV-Zoo framework. The tasks encompass a wide range of embodied challenges, including navigation, object manipulation, and multi-stage goal execution. Each task unfolds as a multi-step trajectory, pairing first-person visual observations with high-level instructions, grounded actions, and natural language rationales that express the agent's intent at every step. Using EmRACE-3K, we establish a benchmark to evaluate the embodied reasoning capabilities of VLMs across three key dimensions: Exploration, Dynamic Spatial-Semantic Reasoning, and Multi-stage Goal Execution. In zero-shot settings, all models achieve success rates below 20%, underscoring the challenge posed by our benchmark and the current limitations of VLMs in interactive environments. To demonstrate the utility of EmRACE-3K, we further fine-tune Qwen2.5-VL-7B using supervised learning followed by reinforcement learning. This approach yields substantial improvements across all three challenge categories, highlighting the dataset's effectiveness in enabling the development of embodied reasoning capabilities.
title EmbRACE-3K: Embodied Reasoning and Action in Complex Environments
topic Computer Vision and Pattern Recognition
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2507.10548