AirGroundBench: Probing Spatial Intelligence in Multimodal Large Models under Heterogeneous Multi-View Embodied Collaboration

A new diagnostic benchmark called AirGroundBench aims to systematically evaluate how well multimodal large language models handle spatial reasoning across heterogeneous air-ground robot views, a capability researchers say remains a persistent bottleneck for embodied intelligence systems [1]. The benchmark, detailed in a paper submitted to the arXiv preprint repository on June 26, 2026, addresses what its creators describe as a gap in existing evaluation frameworks, which have largely focused on single-agent, single-view perception [1]. AirGroundBench is constructed from 11 high-fidelity simulated environments containing 1,021 synchronized air-ground observation pairs, generating roughly 62,000 dual-view visual question answering instances and 115 closed-loop vision-language navigation episodes [1]. The tasks are organized into four capability dimensions of increasing difficulty: spatial perception, cross-view alignment, spatial transformation and reasoning, and embodied decision-making [1]. Evaluations of 13 representative multimodal large language models (MLLMs) under UAV-only, UGV-only, and dual-view input settings revealed consistent performance patterns [1]. Models performed relatively well on basic spatial perception tasks but showed significant difficulty with cross-view alignment and transformation-intensive reasoning, deficits that subsequently impaired sequential decision-making in navigation scenarios [1]. While providing dual-view inputs yielded measurable improvements over single-view variants, a persistent gap from human performance remained, underscoring geometric consistency as a key limitation of current embodied MLLMs [1]. The paper appears on arXiv, an open-access repository that hosts electronic preprints across physics, mathematics, computer science, and related fields and has grown to a submission rate of about 24,000 articles per month as of late 2024 [5]. arXiv papers are moderated but not peer-reviewed before posting [5]. The platform also supports arXivLabs, a framework launched in 2020 that enables community collaborators to develop experimental tools integrated directly on article pages, such as citation explorers and code-finding services [4]. These tools operate under guidelines requiring adherence to openness, community, excellence, and user data privacy, with collaborators granted only minimal and anonymized user data [4].