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.
Figures & tables
Figure 1: Kimodo-Policy as an Instantiation of Kinematic Action Models (KAMs). Top. Kimodo-Policy, best baseline, and second-best baseline. Our policy achieves higher mean success than the compared baselines on the displayed task sets. Middle. Kimodo-Policy instantiates KAMs by inheriting whole-body motion priors from a pretrained motion generator. The compact policy (678M parameters, 92M trainable) combines large-scale loco-manipulation pretraining with a three-stage root–body–hand architecture. Bottom. Observed motion history conditions future action-chunk generation for closed-loop control, as demonstrated in real-world object transfer.
Figure 2: Pretrained Backbones for Humanoid Action Generation. From left to right: VLA, WAM, and KAM paradigms built around vision–language, video generation, and motion generation models. The schematic highlights their primary sources of generative priors: KAMs adapt a pretrained motion generator to produce kinematic action sequences for a downstream tracker, directly transferring learned movement structure into the policy.
Figure 3: Overview of Kimodo-Policy. The colored vertical blocks denote “frozen Kimodo-Root and Kimodo-Body denoisers,” “trainable ControlNet-Root and ControlNet-Body branches injecting DINOv3 ( Siméoni et al., 2025 ) features,” and “trainable HandNet denoiser for bilateral hand control.” Dual-Path Fusion combines motion-query cross-attention with an MLP-based mapping from spatial visual tokens to the future horizon. The predicted root conditions body generation, while future body features condition HandNet; the Motion Projector converts root–body predictions into tracker references. Snowflakes and flames mark frozen and trainable modules, respectively, and dashed boxes indicate Dual-Path Fusion.
Figure 4: From Language-Aligned Motion Priors to Closed-Loop Interaction. Top: motion generated by the original Kimodo illustrates language-aligned whole-body priors. Bottom: real-world object transfer with Kimodo-Policy across successive closed-loop updates. The time markers and history/prediction brackets are schematic and do not specify single-query durations; actual horizons are detailed in Sec. 4.2 . Actual execution repeatedly applies a chunk prefix and replans from updated visual observations and measured motion history.
Figure 5: Composition of the Policy Pretraining Corpus. The corpus combines HIW-500, 14 task subsets from UniFoLM-WBT available when the corpus was assembled, and Humanoid Everyday. UniFoLM-WBT entries released after corpus construction are excluded.
Figure 6: Real-World Teleoperation and Deployment Setup. Left: real-time action-chunk execution, where each policy update supplies a 15-frame prefix at 30 Hz to the 50 Hz tracker, with RTC overlap between consecutive chunks. Right: PICO-based whole-body teleoperation and the humanoid platform equipped with a RealSense D435i camera and Dex3-1 hands.
Figure 7: Motion Profiles of the Collected Real-World Demonstrations. One episode per task is selected to illustrate pronounced task-related motion. Curves show mean angles over six joints per leg or seven per arm. Blue and green denote left and right limbs; solid and dashed curves indicate measured states and recorded motion references, respectively. Task-relevant limbs are emphasized, with other limbs faded. Shaded regions mark the labeled action intervals, and vertical dashed lines delimit their approximate boundaries.
Figure 8: Visualization of the Seven HumanoidArena Tasks. The benchmark includes Football, DoubleDesk, P&Pbox, OpenDoor, SitSofa, Boxing, and VisNavi, covering language-conditioned humanoid behavior across human–object and human–scene interaction with coordinated locomotion, balance, and manipulation.
Figure 9: Qualitative Visual Responses Across HumanoidArena Tasks. Columns show DoubleDesk, Football, P&Pbox, Boxing, OpenDoor, SitSofa, and VisNavi; rows show π0.5 , Diffusion Policy, and Kimodo-Policy. Kimodo-Policy exhibits localized responses near interaction-relevant regions in the displayed examples.
Figure 10: Evaluated SIMPLE Task Suite. The selected tasks cover object handover and mobile pick-and-place, locomotion-integrated object pickup, articulated-object interaction, and household-object manipulation. Panels show (a) Handover, (b) Mobile P&P, (c) XMoveBendPick, (d) CloseDoor, (e) OpenOven, (f) OpenFaucet, (g) Push Office Chair, and (h) OpenTrashCan.
Figure 11: Examples of Real-World Humanoid Manipulation. The image shows representative interaction scenarios involving articulated objects, object transport, and coordinated whole-body motion.
Figure 12: Real-World Task Success. Success rates for GR00T N1.6, Ψ0 , and Kimodo-Policy with and without additional robot-interaction policy pretraining. Both Kimodo-Policy variants retain the pretrained motion backbone. Each task is evaluated for ten trials; the rightmost group reports the mean across eight tasks. The plot label Scratch denotes the variant without additional policy pretraining, while Pretrained denotes the variant with it.
Figure 13: Scaling Policy Adaptation Around a Fixed Motion Backbone. Overall success on the seven HumanoidArena tasks as a function of robot-interaction policy-pretraining data (left) and trainable policy parameters (right). The motion backbone is held fixed in both sweeps. Each panel follows its respective controlled adaptation setting and should not be interpreted as a joint data–capacity sweep.
Figure 14: Policy Adaptation Dynamics and Ablations. Left: downstream training-loss trajectories with and without robot-interaction policy pretraining; both retain the pretrained motion backbone. Right: mean SR on the three HOI tasks after joint seven-task training. Relative to the reference without policy pretraining (light blue), gray bars show random initialization and training of the same Kimodo network (w/o T2M), fine-tuning the pretrained backbone (w/o Frozen), the original Kimodo motion objective (w/o MSE), and removal of root–body gradient detachment (w/o Detach). Dark blue adds robot-interaction policy pretraining. The ablation scores are HOI averages, not seven-task overall SR.
IIIS, Tsinghua University, Beijing, China. · Xiong’an Institute of Artificial Intelligence, Xiong’an, China. · The University of Melbourne, Melbourne, Australia.