Event
TouchSafeBench: A Benchmark for VLM Collision Grounding Released
1 reports · 1 sources
Jun Wang, Xiaohao Xu, and Xiaonan Huang proposed and released TouchSafeBench, a physically grounded benchmark built in Habitat 3.0 for evaluating the collision grounding capability of vision-language models (VLMs). The benchmark contains 2,940 simulated indoor co-presence episodes, covering two types of scenarios—social navigation and social rearrangement—and provides synchronized multi-view RGB-D observations, top-down trajectory maps, calibrated camera metadata, and simulator contact labels. The benchmark examines two deployment-oriented tasks: judging the current safety state, and warning of impending collisions before contact occurs. The authors report that across three frontier or robot-oriented VLMs and nine visual representations, existing models remain far from reliable: the best average Macro-F1 is below 50%, explicit depth information does not automatically translate into collision evidence for the robot body, and robot–scene contact is consistently harder to judge than human contact risk.
Report timeline
- TouchSafeBench: A Human-Robot Collaboration Benchmark for Evaluating Collision Grounding in Vision-Language Models
The authors propose TouchSafeBench, a physically grounded benchmark built in Habitat 3.0 for evaluating the collision grounding capability of vision-language models (VLMs). It contains 2,940 simulated indoor co-presence episodes covering two scenario types, social navigation and social rearrangement, and provides synchronized multi-view RGB-D observations, top-down trajectory maps, calibrated camera metadata, and simulator contact labels.
arXiv cs.RO ·