Vision-language models have achieved impressive progress, yet they still struggle with spatial intelligence–understanding where objects are, how they relate, and how space changes across viewpoints. This limitation matters for embodied AI, autonomous driving, and spatially consistent generation. Meanwhile, rapid advances in spatially enhanced VLMs have produced a scattered literature with inconsistent terminology, methods, and evaluation practices. In this survey, we provide a comprehensive and unified overview of recent advances in spatial intelligence for VLMs. We summarize core concepts behind spatial reasoning in VLMs, analyze why spatial failures occur, and organize existing solutions into a clear framework spanning prompting-based techniques, model improvements, explicit 2D cues, 3D enrichment, and data-driven strategies. We also examine how spatial ability is currently measured and report an empirical study across 37 models and 9 representative benchmarks. Our analysis highlights current best-performing approaches, clarifies when different strategies help or fail, shows the existence of performance gaps across different evaluation datasets and reveals the potential design biases in current spatial understanding benchmarks. By consolidating evidence and outlining open challenges, this survey offers a practical roadmap for building more spatially capable VLMs. We release our
evaluation code
and maintain a curated
paper repository
to support the rapidly growing research on spatial intelligence in vision-language models.
Disheng Liu, Tuo Liang, Zhe Hu et al.· Artificial Intelligence Revi...· 6 citations
Efficiently retrieving relevant clips from large-scale driving logs is essential for data curation, model development, and safety analysis. Structured and rule-based retrieval systems can explicitly target driving events, but typically require expert-defined rules, auxiliary data, and multi-stage perception pipelines. Multimodal embedding models offer a simpler and more efficient alternative by representing each video with a single searchable vector. However, general-purpose models often rely on shortcuts from static scene context and struggle to distinguish motion-centric events, such as turning left versus right or accelerating versus decelerating. In this work, we study how to adapt a general-purpose multimodal embedding model to driving-video retrieval. We first fine-tune Qwen3-VL-Embedding on paired clips and reasoning traces from nuReasoning using an InfoNCE objective. While this stage substantially improves overall retrieval, caption supervision alone remains insufficient for fine-grained motion understanding. We therefore introduce TraVEL (Trajectory-Guided Video Embedding Learning), a motion-aware fine-tuning framework that uses ego-trajectory similarity as a reward within Group Relative Policy Optimization. Trajectories serve only as privileged training supervision; retrieval still operates on single-vector video embeddings without ego poses, expert rules, or auxiliary perception outputs. We further construct a driving-video retrieval benchmark from nuReasoning. Experiments show that TraVEL improves motion-centric retrieval across model scales: relative to SFT, it raises longitudinal and lateral mAP by 9.8 and 4.7 points at 2B, with corresponding gains of 7.2 and 1.5 points at 8B. TraVEL thus combines physically grounded supervision with efficient embedding-based search.
Yi-Chung Chen, P. Jacobson, Tom Lampo et al.· 0 citations