RynnWorld-Latent preprint reports gains from synthetic robot demonstrations

Researchers report that synthetic demonstrations from RynnWorld-Latent improved real-world robot policy success rates over training only on real demonstrations in their evaluated settings. The preprint describes a video world model trained across human videos, robot demonstrations and simulation using a shared representation of actions. The model was pretrained on 23 datasets comprising approximately 553.5M frames, then adapted to target robots. One route generates demonstrations from target robot commands; another transfers motion to an edited starting image and pairs the generated sequence with retargeted robot actions. Experiments included two real-robot platforms. Synthetic training also improved policy success rates on LIBERO and LIBERO-Plus, although the abstract gives no numerical success-rate comparisons.