Robotics Breakthroughs: Tactile Models, World Models, and Data Efficiency

Tactile-native models, compact latent actions, predictive world models, robot-free data, and adaptive reward programs are broadening contact-rich manipulation and sim-to-real learning. N_0-TWAM/N_0-VTLA, HiFi-UMI, predictive navigation, InterEvolve, and 4Director are promising, while SceneActBench and Apple-π reveal persistent physical-reasoning and 3D-interaction gaps.