Research paper
Semantic-Metric Bayesian Risk Fields: Learning Robot Safety from Human Videos with a VLM Prior
A vision-language prior and video-trained likelihood form a Bayesian, spatially varying risk map for robot planning. Demonstrations examine alignment with human preferences and downstream planning uses. Learned relative risk is context-dependent and should not be treated as calibrated injury probability or a validated protective limit.
Reviewed 2026-11-05 · Source and bibliography checked
- Original title
- Semantic-Metric Bayesian Risk Fields: Learning Robot Safety from Human Videos with a VLM Prior
- Publisher
- arXiv
- Authors
- Chen, Timothy, Dominguez-Kuhne, Marcus, Swann, Aiden, Liu, Xu, Schwager, Mac
- DOI
- 10.48550/arxiv.2512.08233
- First archive posting
- 2025-12-09
- Source language
- en
- Source country
- US
- Edition
- v1
- Access
- Open archive metadata and manuscript
- Publication status
- Preprint
Related resources and engineering notes
Source overview
A vision-language prior and video-trained likelihood form a Bayesian, spatially varying risk map for robot planning. Demonstrations examine alignment with human preferences and downstream planning uses. Learned relative risk is context-dependent and should not be treated as calibrated injury probability or a validated protective limit.
Use and review
Original summary of the primary archive abstract and bibliographic record. Specialist technical and native-language review remain pending. The reported method does not establish a deployed safety function or achieved PL.
Sources & editions
Semantic-Metric Bayesian Risk Fields: Learning Robot Safety from Human Videos with a VLM Prior ↗
arXiv · v1
arXiv abstract, authors, submission history and DOI links; first posting 2025-12-09, abstract reviewed in v1. Publisher dates are separate.