MIT's HardFlow enforces hard constraints on generative AI without retraining
MIT News AI · rss · 2026-09-14
MIT researchers led by Navid Azizan have developed HardFlow, a plug-and-play algorithm (published in IEEE TPAMI) that lets pretrained generative models satisfy hard safety, physical, and task-specific constraints at output time without sacrificing quality.
Key idea
- Conventional projection-based sampling forces constraints at every intermediate generation step, limiting exploration and degrading solution quality.
- HardFlow enforces constraints only on the final output, reformulating constrained sampling as a trajectory-optimization problem solved with optimal-control tools. Leveraging the structure of flow-matching models, it decomposes the problem into efficient single-step subproblems solvable at deployment time, with no retraining needed.
Results
- Across robotic manipulation, maze navigation, and text-guided image editing, HardFlow achieved 100% constraint satisfaction while finding better solutions than baselines — e.g., collision-free robotic paths that are also faster to the target, where other methods collided or took far longer.
- Runtime was comparable to or lower than competing methods, making it practical for safety-critical deployment of existing pretrained models.
More from Embodied
- Main approaches for finetuning e2e driving models in close(ish)-loop — abursuc · 2026-09-15
- Unitree upgrades G1 humanoid after 18,000+ units sold, G1+ adds torque and sensing — CyberRobooo · 2026-09-15
- Light Origins' humanoid parkour policy picks walk, vault or climb with onboard sensing only — micoolcho · 2026-09-15
- Agility Robotics unveils Digit 5 humanoid that works alongside people, fence-free — The Decoder · 2026-09-15
- AgentSTAR uses agents for articulated object tracking; Malik hails end of "grad student descent" — CSProfKGD · 2026-09-15
- China opens world's first plant mass-producing a humanoid robot every 10 minutes — mariolefebvre · 2026-09-15