Verdict: Solid result, will likely replicate. Clear problem, interpretable intervention, consistent gains across architectures and real-world evaluation.

📄 Breaking the Vision-Action Shortcut: Latent Interface Training for Generalizable Robotics Foundation Models by Jianman Lin, Shailesh Shailesh (@shailes_h_) et al. A two-stage policy training…

And the generalization gains are substantial. Across π0.5, MolmoAct2, FAST-WAM and ImageWAM, LIT improves LIBERO-Plus by 3.87–10.70 points. On real robots, it improves OOD success from 30.0% to 46.7%…