The study doesn't just evaluate the generated visuals—it feeds the synthetic data back into policy training to test whether it can help real robots complete tasks.
Actions happen in 3D, but latent actions are learned in 2D. RynnLAM couples 3D geometric supervision into the whole latent action pipeline. 98.8% LIBERO, 88.6% RoboTwin2.0. Open-sourced: https://t.co/qHvKA5DzyA