TL;DR
Kimodo-Policy introduces Kinematic Action Models (KAMs), a new paradigm that uses a pretrained motion generator as the foundation for humanoid robot control. By leveraging a frozen text-to-motion backbone and lightweight visual adapters, the model achieves state-of-the-art performance on humanoid loco-manipulation benchmarks and successfully transfers to complex real-world tasks with high efficiency.
What Problem Does This Paper Address?
Humanoid robots must coordinate high-dimensional whole-body movements—balancing their "root" (base), moving their limbs, and manipulating objects—all at once. Current approaches generally fall into two camps: Vision-Language-Action (VLA) models, which excel at semantic understanding, and World Action Models (WAMs), which predict visual dynamics.
However, the authors argue that these models often lack an inherent understanding of the structural regularities of physical motion. Translating a visual plan into a coordinated, executable kinematic sequence remains a massive hurdle. The central question is: Can the structure of whole-body motion itself serve as a foundation for generative pretraining, rather than just being the final output of a policy?
Methods
The authors propose Kinematic Action Models (KAMs), which treat motion generation as the primary backbone. Their specific instantiation, Kimodo-Policy, uses a three-stage architecture:
- Frozen Backbone: It utilizes a frozen Kimodo-G1 text-to-motion model. This backbone already "knows" how humanoids move based on large-scale motion capture data.
- ControlNet-based Conditioning: To make the frozen model responsive to the environment, they add a trainable ControlNet adapter. This branch processes egocentric visual features (from DINOv3) and injects them into the backbone.
- Root–Body–Hand Hierarchy: The model first predicts the global root trajectory, then generates body joint motion conditioned on that root, and finally uses a dedicated HandNet to predict bilateral hand controls.
- Closed-Loop Execution: Instead of just "playing back" a motion, the policy uses motion-history conditioning and action-chunking. It looks at the last 100 frames of history to predict the next 50 frames, executing only a small prefix before replanning.
Datasets
The model was trained and evaluated using a massive breadth of data:
- Pretraining Corpus: A 419-hour dataset (47.6M frames) compiled from HIW-500, UniFoLM-WBT, and Humanoid Everyday. This covers everything from cleaning a living room to picking up medicine.
- Simulation Benchmarks: Evaluated on HumanoidArena (7 leg-critical tasks) and SIMPLE (8 contact-rich tasks like opening faucets and pushing chairs).
- Real-World Data: The team collected 25–30 demonstrations for 8 real-world tasks (e.g., "Football Scoring," "Cushion Sitting," "Box Carrying") using a PICO 4 Ultra teleoperation setup.
Key Insights and Ideas
- Motion as a Prior: The "aha moment" is that a model pretrained to generate human-like motion provides a better starting point for robotics than a model pretrained only on images or text. The motion backbone provides a "movement repertoire" that the policy just needs to ground in the current scene.
- Efficiency through Freezing: By keeping the 678M parameter backbone frozen and only training 92M parameters (the adapters and HandNet), the authors show that you can achieve high performance without the massive computational cost of full-model fine-tuning.
- Hierarchy is Essential: Separating the root (global position) from the body (joint angles) prevents the model from getting overwhelmed by high-dimensional noise, leading to much lower fall rates (AFR).
- Paradigm Shift: Figure 2 highlights how KAMs differ from VLAs and WAMs by making motion the direct target of pretraining.
Key Results
- HumanoidArena: Kimodo-Policy achieved a 72.14% success rate in multi-task settings, outperforming strong baselines like GR00T N1.7 (69.70%) and GigaWorld-Policy (68.39%).
- Stability: It maintained the lowest average fall rate (approx. 4.3%), proving that the motion priors contribute significantly to balance.
- Real-World Transfer: The model achieved a 60% mean success rate across 8 diverse real-world tasks, a 15-percentage-point gain over versions without robot-interaction pretraining.
- Scaling: Performance improved consistently as both robot-interaction data and trainable parameter counts were increased, though with diminishing returns at the highest capacities.
Limitations and Future Work
The authors identify a few key areas for improvement:
- Tracker Dependency: The policy is currently tightly coupled with the SONIC low-level tracker. It is unclear how well it would perform with different control stacks.
- Semantic Depth: While the motion is fluid, the model could benefit from being combined with richer semantic representations (like VLMs) for better high-level task planning in unpredictable environments.
- Embodiment Gap: Future work could explore transferring priors from human motion generators to robot hardware more directly, accounting for the differences in joint limits and torque capabilities.
Summary updated