New method enables AI for safety-critical situations

MIT researchers have devised a method that enables generative artificial intelligence models to generate solutions for high-stakes problems while adhering to strict safety and task-specific requirements. Unlike conventional approaches that enforce constraints at every step, the new technique allows the model more freedom during the generation process and then enforces the hard constraints only on the final output.

This adaptable approach works at deployment time, making it applicable to pretrained generative models without the need for retraining. The method consistently satisfied safety requirements while identifying better solutions than existing techniques in various experiments. The primary innovation, called HardFlow, reformulates hard-constrained sampling as a trajectory-optimization problem using optimal control tools.

By decomposing the problem into smaller subproblems and applying systematic transformations, the researchers developed an efficient algorithm that can be solved at deployment time. This optimization approach enables HardFlow to incorporate additional goals, such as finding the shortest path while avoiding collisions. In tests across robotic manipulation, maze navigation, and text-guided image editing, HardFlow achieved flawless constraint satisfaction while outperforming baseline methods in solution quality.

For instance, it enabled a robotic manipulator to find the quickest path to a target object while avoiding obstacles, a task that other methods either failed to accomplish or achieved suboptimally. The researchers believe that HardFlow can be extended to settings where the AI model itself needs to adhere to strict requirements.