SafeBranch: Branch-Pair Safety Alignment for Embodied Agents

TL;DR - SafeBranch trains embodied vision-language agents to avoid unsafe actions using tightly controlled branch pairs that differ only at a safety-critical step. It improves safety without requiring a critic during deployment or sacrificing task success.

• Builds training pairs by rolling unsafe trajectories back to the violation-causing step and generating a safe alternative action.
• Isolates the safety signal more precisely than imitating safe trajectories or contrasting unrelated safe and unsafe rollouts.
• Evaluations cover IS-Bench, SafetyALFRED, and out-of-distribution settings with unseen tasks and objects.
• Achieves roughly 10× more safe successes than the untrained baseline on the unseen-object variant.