Vision-language-action models (VLAs) have shown promise in robot manipulation, but they often struggle when tasks require precise physical contact. A new arXiv paper from researchers including Carlota Parés-Morlans and Jeannette Bohg investigates why these failures happen and how to fix them.
The authors identify two distinct failure modes. Precision failures occur when the policy cannot produce the small, accurate corrective actions needed during contact. Force failures happen when the policy does not respond appropriately to contact forces. Prior work mostly tried to address contact failures with force-augmented architectures and training-time regularizers, without clearly separating these underlying causes.
To address both, the team proposes FACT, a post-training method that increases low-noise sampling to improve fine action prediction and adds structured force conditioning, guiding action generation with contact state and force history. In evaluation spanning almost 2,500 real-world rollouts across five contact-rich tasks, FACT achieved an average success rate of 66%, compared with 41% for the best prior baseline. The paper was accepted as a Spotlight at CoRL 2026.