What separates a humanoid robot that stumbles through a factory floor from one that fluidly navigates it? According to a substantial new paper published on arXiv by researchers at the Chinese University of Hong Kong, Shenzhen, and the Shanghai Artificial Intelligence Laboratory, the answer lies in three carefully coordinated decisions: how the robot learns, what data it trains on, and the architecture that processes both. The paper, titled "Scaling Behavior Foundation Model for Humanoid Robots" and submitted on July 16, 2026, demonstrates that getting these three factors right reduces the Mean Per-Keypoint Position Error (MPKPE) by over 10% in local mode and an astounding 82% in global mode compared to existing humanoid controllers.
The Problem with Today's Humanoid Controllers
Humanoid robots have made dramatic headlines over the past year, with companies like Tesla, Agility Robotics, and Figure deploying prototypes in factory settings. But the control software behind these machines remains a bottleneck. Existing approaches typically tackle isolated sub-problems: walking gaits here, arm manipulation there, balance recovery somewhere else. This fragmented approach produces robots that can walk in a straight line or pick up a box but struggle when asked to combine movements into fluid, whole-body coordination. The paper identifies this as the central challenge: humanoid control requires natural whole-body coordination, precise real-time responses to control signals, and robust generalization across diverse environmental contexts. Behavior Foundation Models, or BFMs, have emerged as a promising solution by training on large-scale behavioral data, but until now the field lacked a clear recipe for scaling them effectively.
The Three-Part Scaling Recipe
The researchers propose that effective scaling of BFMs requires coordination across three distinct components, each addressing a different layer of the control problem. The first is the learning paradigm itself. Instead of treating different robot behaviors as separate problems to solve, the authors reformulate diverse humanoid control tasks as a single problem: the reproduction of integrated whole-body behaviors in a global reference frame. This motion tracking paradigm means the robot learns to reproduce demonstrated behaviors in the real world, not just in simulation, creating a direct bridge between training data and physical deployment.
The second component addresses the data strategy. The paper demonstrates that the strategic synergy between on-policy rollout quantity and reference motion diversity drives performance more than either factor alone. In plain terms: you need both enough real-time interaction data generated by the robot itself and a sufficiently diverse library of reference motions for the robot to learn from. Skimping on either produces measurably worse results. This finding has direct implications for anyone building a humanoid training pipeline, because it suggests that data collection and simulation infrastructure must be designed together rather than optimized in isolation.
The third component is the model architecture itself. The researchers introduce the Humanoid Transformer, an expressive and scalable architecture designed specifically for whole-body control. The Humanoid Transformer facilitates the natural emergence of structured behavioral representations, meaning the model learns to organize its internal representations around meaningful physical concepts such as limb position, joint torque, and balance state without being explicitly programmed to do so. This architectural choice is what enables the dramatic improvement in global mode, where the model must coordinate every joint simultaneously rather than controlling localized limb movements independently.
Benchmark Results and Real-World Deployment
The paper reports extensive experiments in both simulation and real-world deployment. In local mode, where the model controls specific limb movements, the Humanoid Transformer reduces MPKPE by over 10% compared to existing controllers. In global mode, where the model controls the full body as an integrated system, MPKPE drops by 82%. That is not an incremental improvement. It is the difference between a robot that looks like it is operating its limbs versus one that moves the way a human does. The researchers attribute this gap to the structured behavioral representations that emerge naturally from the Humanoid Transformer architecture, which allows the model to reason about the body as a connected whole rather than a collection of independent actuators.
Real-world deployment results confirm that the simulation findings transfer to physical robots, though the paper notes that the sim-to-real gap remains a challenge that requires additional fine-tuning for production environments. The authors release their work under the arXiv identifier 2607.15163, and the paper is available in full at https://arxiv.org/abs/2607.15163.
What This Means for Builders
For founders and engineers building in the humanoid robotics space, this paper offers several actionable takeaways. First, the three-component scaling recipe provides a framework for evaluating your own training pipeline. If your robot is underperforming, the problem may not be your hardware or your reward function. It may be that your learning paradigm, data strategy, and architecture are misaligned. Second, the 82% improvement in global mode suggests that whole-body coordination is where the next wave of robotic capability will come from. Founders building for specific verticals like warehouse logistics or eldercare should prioritize global coordination over localized limb control, because the integrated system produces far more natural and reliable movement. Third, the emphasis on on-policy rollout data combined with diverse reference motions means that simulation infrastructure and data collection are not separate concerns. If you are building a humanoid robotics startup, your simulator and your motion capture pipeline need to be designed from day one as two halves of a single system.