WanderDream: A Benchmark for Situated Spatial Reasoning from Panoramic Views

WanderDream is a benchmark dataset for evaluating situated spatial reasoning without active exploration. It was created by author lrp123 and is hosted on Hugging Face, with a last recorded update on 2026-06-25. The dataset challenges models to reason about future visual scenes and obstacles based on a starting panoramic view and a natural-language target description.

Use Cases

• Benchmarking visual reasoning models based on panoramic starting views.
• Training world models for emulative simulation based on target situation descriptions.
• Evaluating AI's ability to predict visual obstacles based on natural-language instructions.

Strengths

• Dataset is associated with a published research paper titled 'What if? Emulative Simulation with World Models for Situated Reasoning'.
• The core challenge is clearly defined: reasoning from a panoramic starting view and a natural-language target description.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Freshness

Last updated 2026-06-25 22:34:50; freshness should be verified.