# Robot Data Synthesis

RoboTryOn treats a first-person human manipulation video like a virtual try-on: it keeps the demonstrated motion and swaps the person for a target robot, then an inverse dynamics model with depth features reads joint action labels from the synthesized video. On two real humanoid robots with dexterous hands, policies trained only on synthesized demonstrations raise average success on unseen tasks from 2.1% to 60.0% and reach 45.9% on bimanual tasks even though every teleoperated demonstration is single-arm; training on translated human video alone matches teleoperated-only success at about a third of the collection cost.

RoboTryOn treats a first-person human manipulation video like a virtual try-on: it keeps the demonstrated motion and swaps the person for a target robot, then an inverse dynamics model with depth features reads joint action labels from the synthesized video. On two real humanoid robots with dexterous hands, policies trained only on synthesized demonstrations raise average success on unseen tasks from 2.1% to 60.0% and reach 45.9% on bimanual tasks even though every teleoperated demonstration is single-arm; training on translated human video alone matches teleoperated-only success at about a third of the collection cost.

RynnWorld-Latent conditions a video world model on continuous latent actions extracted by a frozen RynnLAM encoder, giving human, robot and simulation video one shared interface without aligning native control spaces. Pretrained on 23 datasets totaling about 553.5M frames and then post-trained per target robot, it reports better video-prediction metrics on public benchmarks and two real platforms, and data from latent-action synthesis and digital teleoperation raises downstream policy success in the evaluated settings.

RynnWorld-Latent conditions a video world model on continuous latent actions extracted by a frozen RynnLAM encoder, giving human, robot and simulation video one shared interface without aligning native control spaces. Pretrained on 23 datasets totaling about 553.5M frames and then post-trained per target robot, it reports better video-prediction metrics on public benchmarks and two real platforms, and data from latent-action synthesis and digital teleoperation raises downstream policy success in the evaluated settings.