Vision‑language model based embodied agents can follow instructions but often breach safety constraints during execution, a problem framed as interactive safety.
Training safe behavior is hard because safety and task success are distinct objectives and safety only appears at a few critical steps in a trajectory. Standard supervision is insufficient: imitating safe trajectories does not explain why they are safe, and contrasting arbitrary safe and unsafe trajectories mixes safety signals with unrelated differences.
The SafeBranch framework rolls back each unsafe rollout to the safety‑critical step that caused the violation, queries the actor for a safe alternative, and pairs the original action with the alternative so that the two branches differ only at that step. This branch‑pair alignment enables the trained actor to act safely at deployment without any safety critic.
Experiments on IS‑Bench, SafetyALFRED, and out‑of‑distribution variants with unseen tasks and objects show that SafeBranch maintains task success while markedly improving safety success, achieving roughly ten‑fold more safe successes on the unseen‑object variant.
Blogger's Review: The approach cleverly leverages environment rollback to generate aligned data, addressing the sparsity of safety signals and demonstrating a practical path toward safer embodied agents in real‑world settings.