Space & aviationResearch findings
Investigating Physical Safety Awareness in Vision-Language Models
Original published 2026.05.29 21:04 KSTAdded to Blackdeck 2026.11.04 00:38 KST
What has changed?
A physics-grounded benchmark, TouchSafeBench, is introduced to evaluate collision grounding in vision-language models.
Key points
- Analyzes limitations in vision-language models' ability to detect physical collisions.
- Introduces TouchSafeBench to assess grounding between visual observations and physical contact.
- Models fail to accurately predict collisions between robot bodies, scenes, or humans.
Connect it to your work
The study reveals a gap between visual fluency and physical accountability in embodied vision-language models.
What remains unproven
The abstract reports simulation results; real-world performance is not established by this input.
Originals used for this note
Links open the original publisher in a new tab.