KinoGen researchers report 90% contact preservation in 8 robot tasks

KinoGen researchers reported preserving more than 90% of annotated contacts across eight humanoid loco-manipulation tasks in an arXiv preprint submitted on Oct. 14.

The preprint matters because humanoid robots need usable motion references before they can be trained to move objects while walking. The authors said KinoGen also produced less penetration and foot skating than video-based baselines across the eight tasks.

KinoGen is a framework for creating customizable references for humanoid loco-manipulation, a robotics term for moving through space while handling objects. The paper was written by Qi Wu, Lulu Zhao, Sanghyun Hahn, Rishi Yennu, Emmy Yang, Dingqi Zhang, Tao Pang, Jiuguang Wang and Kuan Fang.

The authors said KinoGen uses a pretrained video generator conditioned on calibrated renders of the target robot and object in a simulated task. It then reconstructs the reference by fitting the same robot and object models to generated two-view video, keeping the output within the robot’s own configuration space rather than relying on a human-body proxy.

The paper said users can edit a keyframe to specify a target object pose, regenerate video around that state and obtain a reference that meets the prescribed pose. The authors said customized references satisfied every prescribed target and produced substantially less penetration than direct interpolation.

The work remains early research. The arXiv paper is a preprint and has not been peer reviewed, and the source text does not provide commercial performance data, training cost, sample sizes beyond the eight tasks, or a comparison against deployed industrial systems.

The authors said tracking policies trained on the generated references were deployed on a Unitree G1 humanoid with dexterous hands. The paper lists a project webpage for the work.

Source: arXiv preprint, Oct. 14, 2026