Teaching robots two hands and ten fingers at once

Controlling robots with five fingers per hand is harder than simple gripper manipulation, but existing vision-language-action models skip it. Dexora combines a custom exoskeleton for arm control and Apple Vision Pro for finger tracking to collect 10K teleoperated episodes, then trains a diffusion-transformer policy that learns which demonstrations are trustworthy via an offline discriminator. The result: 66.7% success on dexterous manipulation—25% better than baseline VLAs—plus code, models, and datasets released openly.