DIGIT is a low-cost, compact, vision-based tactile sensor specifically engineered for robotic in-hand manipulation tasks. Developed by researchers at Meta AI and NYU, the sensor consists of a soft gel fingertip cover deformed by contact forces, a small camera underneath capturing the deformation field, and a thin elastomer membrane that creates a high-contrast pattern when pushed. By analyzing images of the interior finger surface, the system reconstructs the spatial distribution of contact forces and local geometry at sub-millimeter resolution. The design philosophy behind DIGIT prioritizes manufacturability and affordability. Each sensor costs approximately $25 in parts, can be assembled with off-the-shelf cameras and 3D-printed molds, and is small enough (15 mm diameter) to be mounted on standard robotic fingertips. Unlike previous vision-based tactile sensors that required bulky optics or complicated mounting jigs, DIGIT achieves miniaturization by using a wide-angle lens placed extremely close to the deformable gel surface, capturing roughly 640-by-480 pixel images at over 60 Hz. In evaluation, DIGIT demonstrates superior performance in grasp stability prediction, slip detection, and object surface-texture classification. When mounted on a dexterous hand, it enables precise force-controlled object reorientation, insertion, and surface following. The sensor’s output lends itself naturally to learning-based approaches: recent studies feed raw DIGIT images directly into convolutional and transformer-based policy networks, bypassing explicit force reconstruction and learning manipulation strategies from raw tactile pixels. DIGIT has become the de facto open-source tactile sensor in manipulation research. The hardware design files, embedded firmware, and driver code are publicly released, and a growing ecosystem of compatible robotic hands (including the Meta/MyoHand and several Franka-compatible grippers) has emerged. For researchers working on contact-rich tasks, adding DIGIT provides rich local contact information that is either invisible or noisy when relying solely on wrist force/torque sensing.