Humanoids can learn new skills by tracking high-quality reference...
Qi Wu@Qi_Wu577
8 views
Oct 16, 2026
~2 min read
Advertisement
1
Humanoids can learn new skills by tracking high-quality reference motions, which today come from humans through motion capture or video. Routing every reference through a human proxy limits scalability and compounds reconstruction errors with the embodiment gap. KinoGen synthesizes loco-manipulation references directly for the target robot through kinematically grounded video generation. 🧵 (1/7)
0:25
2
KinoGen generates multi-view robot videos and directly fits the target robot and object models to them. Compared with video-based methods that rely on human retargeting or robot-specific pose detectors, this avoids errors introduced by transferring motion between embodiments and recovers complete body, hand, and object references for the target robot and scene by construction. (2/7)
0:31
3
A generated reference may still miss prescribed task configuration. Because KinoGen uses the same robot and object models for generation and reconstruction, it can render corrected keyframes back into the generator to regenerate the surrounding motion. Reconstruction then enforces the edits exactly, closing the loop for precise reference customization. (3/7)
0:18
4
KinoGen synthesizes robot and object motion across multiple task instances and interaction modes. (4/7)
0:16
5
Kinematic grounding shows up directly in reference quality. Against three video-based pipelines on eight tasks, KinoGen achieves the lowest penetration and foot skating and preserves over 90% of annotated contacts. Policies trained on these physically consistent references outperform those trained on every baseline, with the highest success rate and the lowest tracking error.(5/7)
0:06
6
Policies trained on KinoGen references are deployed on a Unitree G1 with dexterous hands, where they mop the floor, lift a chair, kneel to pick up a ball, and sit on a stool. Every reference motion originates from a generated video of the robot itself, without human demonstrations, a human body proxy, or retargeting.(6/7)
0:26
7
Project page: kinogen-humanoid.github.io
Paper: arxiv.org/pdf/2610.20552
Huge thanks to the team from Cornell University and @multiplylabs !
@Qi_Wu577 , @LuluZhaoAI , @hihighhahn , Rishi Yennu, Emmy Yang, @realdingqi , @pangtao22 , Jiuguang Wang, @KuanFang
Paper: arxiv.org/pdf/2610.20552
Huge thanks to the team from Cornell University and @multiplylabs !
@Qi_Wu577 , @LuluZhaoAI , @hihighhahn , Rishi Yennu, Emmy Yang, @realdingqi , @pangtao22 , Jiuguang Wang, @KuanFang
