Language-conditioned safety filters cut robot constraint violations, shows arXiv paper

As vision-language-action (VLA) models let a single robot policy handle diverse natural-language tasks, safe deployment still demands adapting to changing safety rules across users and environments. Existing safety filters are constraint-specific, forcing redesign or relearning whenever requirements shift. In a new arXiv paper, researchers Ihab Tabbara, Yuxuan Yang, and Hussein Sibai investigate language-conditioned safety filtering, where a Hamilton-Jacobi safety actor and critic are conditioned directly on language-specified constraints. This allows the filter to interpret constraints like 'don't knock over the cup' without being retrained for each new scenario.

The team evaluated the approach in vision-based settings across pick-and-place, table-wiping, and block-stacking tasks. Results show that language-conditioned filters reduce constraint violations compared to baseline methods, and exhibit partial transfer to unseen constraint instances within the same constraint families. While the transfer isn't fully general, it demonstrates that language can be a viable interface for safety specification in robot learning. This is a step toward generalist robotic systems that can be safely deployed in dynamic environments, where safety requirements are as varied as the tasks themselves.