Large-scale visual pretraining has improved generalist robot policies, but contact-rich dexterous manipulation remains a challenge. The authors of the new paper argue that key physical signals—force, compliance, and incipient slip—are hard to observe from images, leaving a gap that vision-only pretraining cannot fill.
The paper introduces Spark-0, a tactile pretraining approach designed to address this gap. As the title indicates, the method focuses on spatially aligned tactile representations, though the abstract does not detail the underlying architecture or training procedure.
The significance lies in complementing visual pretraining with touch-based information. By targeting the contact signals that images miss, Spark-0 could help robot policies handle tasks where dexterity depends on sensing the physical interaction itself.