Position paper proposes embodied AI safety taxonomy for generalist robots

Majumdar_Ani · x · 2026-09-12

A position paper by Andrea Bajcsy, Anirudha Majumdar et al. (arXiv:2609.06326) argues collision- and force-based safety is insufficient when robots can 'do anything.' New safety notions must account for context (powering down a building is only safe during maintenance), user intent ('clean the kitchen' implies not mixing bleach and ammonia), and hard-to-model physical consequences (burning food). The authors call for an 'embodied AI safety' era, present a taxonomy of emerging risks, and lay out a full-stack research agenda for safe generalist robot deployment.

Original post →

More from Embodied

Embodied channel →