The paper argues that human demonstrations offer a scalable alternative to costly robot data collection for dexterous manipulation. However, most human datasets do not include force measurements, which are critical for tasks that involve contact and fine motor control. The abstract notes that dexterous manipulation requires not only visual and proprioceptive information, though the description cuts off before listing the full set of requirements.
To address this gap, the authors propose FACT, a multimodal force-action-track foundation model. The model is designed to integrate force, action, and track modalities, presumably allowing robots to learn from human demonstrations that include the missing force channel. The name and framing suggest a focus on enabling more robust human-to-robot transfer of dexterous skills.
The source is a brief abstract, so details on architecture, training data, and experimental results are not yet available. Still, the core motivation is clear: without force data, human demonstrations may be insufficient for dexterous manipulation, and FACT aims to fill that gap with a multimodal foundation-model approach.