The authors used three VLM-free FormulaOne driving policies (which never received the score) to generate 180 episodes with 130 isolated contact onsets.
Frozen vision-language safety score tracks scene template, not hazards directly
Controlled experiments show that CLIP-based safety signals in reinforcement learning reflect caption-bank scene similarity rather than actual hazard detection.
Academic
Samuel Tetteh · Cody Fleming
Iowa State University
Research Digest··3 min read
Samuel Tetteh and Cody Fleming from Iowa State University audit a frozen CLIP prompt-margin safety score in a driving RL setup.
Why this paper
From Iowa State University
In one line
A frozen CLIP safety score measures scene resemblance to its caption bank, not hazard detection.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§