Robotic manipulation policies increasingly rely on vision-language-action (VLA) models, but vision alone may not be sufficient. Force feedback provides these policies with physical observations that can guide action selection and execution. However, the paper notes that differences in tasks, demonstrations, and control settings across existing systems hinder direct comparison.

To address this, the authors introduce ForceBench, a benchmark specifically designed for force integration in robotic manipulation. By providing a standardized evaluation framework, ForceBench aims to make progress in this area more measurable and reproducible across different research efforts.