MIT’s HardFlow tests a different route to enforcing AI constraints

A robot route can be short and still hit an obstacle. MIT researchers have developed HardFlow to help generative models satisfy non-negotiable requirements while also improving qualities such as journey time.

The team comprises lead author Zeyang Li, Kaveh Alim and senior author Navid Azizan. Adam Zewe’s MIT News account describes the research, published in IEEE Transactions on Pattern Analysis and Machine Intelligence.

Common projection-based approaches force partial answers to meet constraints throughout generation. HardFlow instead steers that internal process while enforcing the hard requirements on the finished answer. The discarded intermediate samples need not themselves be usable solutions.

How the tests turned out

Using optimal-control methods, the algorithm divides a difficult optimisation problem into smaller steps. It can pursue an additional goal, such as reducing travel time, while keeping the final route clear of obstacles.

The team tested robot manipulation, navigation through mazes and image changes directed by text. They report that no required constraint was breached in those experiments and that the answers scored better than those from comparison methods.

HardFlow also generally needed no more computing time than its rivals, and sometimes less. The findings demonstrate results on those tasks, not a universal safety guarantee.