Verified Events

Xspark AI introduces Spark-0 tactile pretraining model and HumanTouch corpus

• Spatial contact maps and a frozen 3D geometry teacher supervise visual representations during pretraining.
• On matched visual data and compute, reported vision-only Ego Humanoid success increases from about 52.5% to71.7%.
• One checkpoint supports vision-only and tactile-conditioned inference; touch-supervision heads are removed at inference.