RoboVista: 474 Multiple-Choice Questions Over Real Robot Scenes

RoboVista is a benchmark of 474 multiple-choice questions for evaluating vision-language models on robot perception and reasoning skills. The dataset pairs images from real robot deployments with answer choices and ground-truth answers. It was created by sy-xie and last updated on July 6, 2026.

Use Cases

• Benchmarking vision-language model performance on robot perception tasks based on real-world robot images.
• Evaluating model reasoning skills for robotics based on multiple-choice question formats.
• Training or fine-tuning models for scene understanding in diverse robot environments like surgical setups or driving scenes.

Strengths

• Contains 474 multiple-choice questions, providing a quantifiable evaluation scale.
• Uses images from real robot deployments, including wrist cameras, overhead views, driving scenes, and surgical setups.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2026-07-06 22:54:25; freshness should be verified.

Provenance

Source

sy-xie on Hugging Face

Collection Method

Likely compiled from real robot deployments for benchmark creation.

Freshness

Last updated 2026-07-06 22:54:25.