🖐️ Robotics has a tactile data problem. Vision can leverage internet-scale images and videos, but touch is fundamentally different: every tactile sample has to come from a physical interaction. Different objects, contact locations, poses, forces, and different sensor instances can all produce very different tactile observations. Collecting these variations in the real world is slow, expensive, and difficult to scale. ❓So we asked: Can we synthesize diverse and useful tactile interactions from just a single real tactile image? Introducing ControlTac : Scaling Tactile Data with Physically Controlled Tactile Image Generation, accepted by #CoRL2026 🧵👇 ✦ Using a single tactile image with initial contact for reference ✦ Conditioned on contact force and pose for generation ✦ Ensuring realism and diversity in the generated images ✦ Generating images for downstream tasks like force estimation and pose estimation ✦ Evaluate on real-robot tasks like object weighting, insertion, and imitation learning 🌐 Website: dongyuluo.github.io/controlt… 📖 Paper: arxiv.org/abs/2505.20498 📷 Video: youtube.com/watch?v=7WzQ9s1c… 💻 Code: github.com/DongyuLUO/Control… 1/n