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

Yunheng 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

November 2026

TL;DR Kimodo-Policy turns a frozen pretrained text-to-motion generator into a closed-loop humanoid policy by adding visual ControlNet branches and a hand head. It reaches 72.14% on multi-task HumanoidArena and 60% mean success on eight real-world loco-manipulation tasks with 92M trainable parameters.

What they do

Humanoid policies usually start from a vision-language model or a video model, neither of which is pretrained on robot motion itself. This paper proposes Kinematic Action Models: use a pretrained whole-body motion generator as the action backbone. Kimodo-Policy keeps the Kimodo-G1 text-to-motion model frozen and trains small modules around it so that it reacts to the camera image and the robot’s own executed motion.

How it works

Visual conditioning. Trainable ControlNet branches feed frozen DINOv3 features of the egocentric image into the frozen root and body denoisers, through cross-attention plus an MLP path, with zero-initialized projections.
Root, body, hand. The root stage predicts translation and heading, the body stage predicts joint motion given the root, and a separate diffusion head (HandNet) predicts open/close commands for both hands. Gradients are detached between stages.
Closed loop. Each query is motion completion: up to 100 frames of measured motion history are fixed and 50 future frames (30 Hz) are generated; 15 frames are executed by the SONIC whole-body tracker before replanning.
Pretraining. The trainable modules are pretrained on 419 hours of public humanoid demonstrations. 678M parameters in total, 92M trainable.

Results

HumanoidArena, one policy for seven tasks: 72.14% success, against 69.70% for GR00T N1.7. It leads on scene-interaction tasks but trails on object-interaction tasks (about 56.1% against 62.22%).
SIMPLE, eight tasks: 75.83% against 75.42% for Psi-0, with baseline numbers taken from other reports. Success drops from 87.50% at the easiest randomization level to 68.75% at the hardest.
Real humanoid, eight tasks, ten trials each: 60% mean success (48/80), against 45% without policy pretraining, 35% for Psi-0 and about 31% for GR00T N1.6.
Ablation on object-interaction tasks: a randomly initialized backbone gets 18.33%, a fine-tuned pretrained one 31.67%, the frozen pretrained one 51.67%.

Limitations

The policy is built and tested only with the SONIC tracker; other low-level controllers are not supported. Policy pretraining lowers single-task HumanoidArena success (67.14% against 69.76%). Real-world rates come from ten trials per task, and the effect of real-time chunking is not ablated.

Figures

From Figure 1: Top, mean success of Kimodo-Policy against the best and second-best baselines. Middle, the policy inhe... — From Figure 1: Top, mean success of Kimodo-Policy against the best and second-best baselines. Middle, the policy inherits whole-body motion priors from a pretrained motion generator (678M parameters, 92M trainable) with a root, body and hand architecture. Bottom, observed motion history conditions the next action chunk during real-world object transfer.

From Figure 3: Frozen Kimodo-Root and Kimodo-Body denoisers, trainable ControlNet-Root and ControlNet-Body branches t... — From Figure 3: Frozen Kimodo-Root and Kimodo-Body denoisers, trainable ControlNet-Root and ControlNet-Body branches that inject DINOv3 features, and a trainable HandNet denoiser for the two hands. The predicted root conditions body generation, future body features condition HandNet, and the root and body predictions become tracker references.

From Figure 11: Examples of real-world humanoid manipulation, with articulated objects, object transport and coordina... — From Figure 11: Examples of real-world humanoid manipulation, with articulated objects, object transport and coordinated whole-body motion.

From Figure 12: Real-world success rates for GR00T N1.6, Psi-0, and Kimodo-Policy with (Pretrained) and without (Scra... — From Figure 12: Real-world success rates for GR00T N1.6, Psi-0, and Kimodo-Policy with (Pretrained) and without (Scratch) robot-interaction policy pretraining. Ten trials per task; the rightmost group is the mean over the eight tasks.

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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong