SafeBranch Enhances Embodied Agent Safety Alignment

Embodied agents, particularly those powered by vision-language models, are capable of executing complex instructions but frequently encounter issues with violating safety constraints. This problem, known as interactive safety, is challenging to address because safety and task completion are distinct objectives, and safety violations often occur at only a few critical points within a longer sequence of actions. Traditional training methods, such as imitating safe trajectories or contrasting arbitrary safe and unsafe ones, often fall short because they either teach behavior without explaining the underlying safety rationale or mix safety signals with irrelevant differences. To overcome these limitations, researchers have developed SafeBranch, a novel framework designed to align an embodied actor specifically on safety. SafeBranch operates by constructing "branch pairs" directly from the agent's own unsafe rollouts. When an unsafe action is detected, the environment is rolled back to the exact safety-critical step that caused the violation. The actor is then prompted to generate a safe alternative action, creating a pair where the original unsafe action is contrasted with the newly generated safe one, with all other aspects of the trajectory remaining identical. This targeted alignment allows the trained actor to act safely during deployment without needing an external critic. Experiments on various benchmarks, including out-of-distribution scenarios with unseen tasks and objects, demonstrate that SafeBranch reliably handles safety concerns without compromising task success. It achieved significantly more safe successes compared to untrained baselines, particularly in challenging unseen-object variants.