lvesucces2027apache-2.0

Safety Inspector ICRA 2027 Held-Out Test Benchmark

A held-out test benchmark containing 50 tabletop scenes and 100 query tasks for evaluating vision-language model safety inspection, rule selection, and test-time adaptation in robotic manipulation.

Downloads0
Episodes100

Why This Matters for Physical AI

This benchmark dataset evaluates vision-language models' ability to inspect robot safety and correct manipulation plans, advancing research in safe autonomous robotic systems that can reason about task constraints and adapt online.

Technical Profile

Modalities
rgblanguage
Environment
labtabletop
Task Types
visual-question-answeringsafety-inspectionmanipulation
Episodes
100
Annotation Types
language_instructionsaction_labelsreward_labels
License
apache-2.0
Part of the Safety Inspector family

Access

Need custom rgb data?

Claru builds purpose-built datasets for lab applications with dense human annotations and quality assurance.

Request a Sample Pack

Related Datasets