NTH

Kimodo-Policy: Learning Kinematic Action Models for Humanoid Loco-Manipulation

AuthorsYunheng Wang, Jiaxi Zhang, Wei Lu, Yuetong Fang, Taowen Wang, Yixiao Feng, Shuning Zhang, Junzhe Xu, Yichi Wang, Zizhao Yuan, Kailun Huang, Lusong Li, Mingqian Zhou, Zecui Zeng, Renjing Xu

AffiliationsHKUST(GZ) · JD Explore Academy · BJUT · Cambridge ★

November 19, 2026 2 min read
Watch on YouTube
The one-line take

Kimodo-Policy turns a pretrained whole-body motion generator into a control policy for humanoid robots that walk and manipulate objects.

Key results

678M
Total model parameters

Kimodo-Policy has 92M trainable parameters.

419
Pretraining corpus

Hours of robot-interaction demonstrations.

72.14%
HumanoidArena multi-task success

Seven-task result; NVIDIA GR00T N1.7 reaches 69.70%.

75.83%
SIMPLE mean success

Across eight tasks, compared with Ψ0 at 75.42%.

60%
Real-world mean success

Across eight physical-robot tasks.

What the paper found

Kimodo-Policy introduces Kinematic Action Models, which use a pretrained motion generator itself as the action-generation backbone for humanoid control. It keeps Kimodo-G1 frozen and adapts its root and body motion with ControlNet visual conditioning from Meta’s DINOv3, then predicts hand commands in a separate stage. The root–body–hand hierarchy combines with motion-history constraints and repeatedly replanned action chunks to connect coherent motion synthesis with closed-loop feedback. Trained on a 419-hour robot-interaction corpus, the 678M-parameter policy has 92M trainable parameters. On seven-task HumanoidArena, it reaches 72.14% success, ahead of NVIDIA GR00T N1.7 at 69.70%; on SIMPLE it records 75.83%, close to Ψ0 at 75.42%. On a physical humanoid, it averages 60% success across eight tasks. Ablations support retaining the pretrained motion backbone, while results also show that interaction-data pretraining helps in multi-task and real-world settings. The evidence is promising but limited to a control stack built around the SONIC tracker; transfer to other trackers remains untested.

Original abstract

Humanoid loco-manipulation requires continuous coordination of high-dimensional whole-body motion to realize task objectives. While vision-language-action models and world action models transfer semantic and visual-dynamics priors to robot control, these priors do not necessarily capture the structure of robot motion. We introduce Kinematic Action Models (KAMs), a paradigm that treats whole-body motion as a direct pretraining target and uses a pretrained motion generator as the primary action-generation backbone. KAMs ground learned temporal and cross-body coordination priors through task conditioning, visual observations, and execution feedback. We present Kimodo-Policy as an initial instantiation, combining a frozen Kimodo text-to-motion backbone with lightweight ControlNet-based visual conditioning. A root-body-hand hierarchy coordinates locomotion and manipulation, while motion-history conditioning and repeated action-chunk prediction bridge motion synthesis and closed-loop control. Following large-scale loco-manipulation pretraining, the policy achieves the highest overall success among compared policies on multi-task HumanoidArena (72.14%), competitive performance on SIMPLE, and 60% mean success across eight real-world tasks, with 678M parameters (92M trainable). Ablations support the contribution of the pretrained motion backbone. Together, these results provide initial evidence for motion generation as a complementary foundation for humanoid action learning.

Read the original paper

More in Embodied AI

Browse all 69 papers →
03Embodied Ai

JustRUA: A Controlled Study of GPT-6 Robot-Use Agents

Zefan Wang, Yufan Dang, Xuanwei Shi, Mingqi Gao, Cheng Yin, SikYuen Tam, Hanyu Liu, Yasheng Wang, Yaxi Lu, Songyang Zhang, Xin Cong, Yuan Yao

A leaner harness helps a GPT-driven robot perform much better in simulation, showing that agent design choices can matter as much as learning.

Read analysis