We’ve been developing tactile sensing hardware and algorithms for quite some time, from sensor design and imitation learning (3D-ViTac) to scaling real-world data (Touch in the Wild). This new work completes one of the most critical pieces of the ecosystem: **Achieving precise sim-real alignment for real-to-sim-to-real transfer of visuo-tactile policies.** The image below is one of my favorites. It shows the same set of robot interaction trajectories in both simulation and the real world, with their tactile signals compared side by side. After converting both into histograms, you can see how closely they align, a level of fidelity that greatly enhances sim-to-real transfer and enables small domain randomization, which is essential for fine-grained manipulation tasks. Congrats to the team on VT-Refine (#CoRL2025)! Website: binghao-huang.github.io/vt_r…