Robbyant Team

“TourPhysics achieves camera trajectory adherence (ATE) of 0.0233 vs. 0.0560 for Sora-2 and 0.1031 for LingBot-Cam”

“The paper argues this is the root cause of why current video world models 'lose physical or spatial consistency over long horizons'”

“LingBot-Cam achieves camera ATE of 0.1031 and object ADE of 0.0833; LingBot-Act achieves 0.1235 and 0.0834.”

“We provide additional experiments for LingBot-World [82] in App. E, an autoregressive world model built on a 14B backbone...”