ForceFlow boosts robot precision by 37% using force-driven flow matching

ForceFlow, a new framework from researchers at multiple institutions, solves one of robotics' toughest challenges: contact-rich manipulation. Unlike existing imitation learning methods that struggle with complex contact dynamics, ForceFlow treats force as a global regulatory signal through an asymmetric multimodal fusion architecture. The system divides tasks into a vision-dominant approach stage (using VLM-based pointing) and a touch-dominant interaction stage, with a Vision-to-Force (V2F) handover mechanism that explicitly decouples spatial generalization from contact regulation.

In experiments across six real-world contact-rich tasks, ForceFlow outperformed the strong baseline ForceVLA by 37% in success rate while maintaining significantly lower cost. The framework also demonstrates accurate force signal prediction, superior contact force self-regulation, and impressive zero-shot out-of-distribution generalization. By integrating flow matching with force-aware reactive control, ForceFlow points toward a more robust and generalizable approach for robots that need to physically interact with the world—from assembly lines to household chores.