KinoGen: Customizable Humanoid Loco-Manipulation References via Kinematically Grounded Video Generation

KinoGen is a framework that synthesizes customizable humanoid loco-manipulation references through kinematically grounded video generation. It conditions a pretrained video generator on calibrated renders of the target robot and object in a simulated task instance, then recovers the reference by fitting the same models to the generated two-view video. The method avoids the need for human body proxies or retargeting, and body, hand, and object priors are used to initialize the reconstruction, which is refined through direct optimization against cross-view evidence. KinoGen outperforms video-based baselines in terms of penetration and foot skating, and preserves over 90% of annotated contacts across eight tasks. Tracking policies trained on the synthesized references are deployed on a Unitree G1 humanoid with dexterous hands.