Weishuai Zeng, Kangning Yin, Xiaojie Niu +9 more
Getting a humanoid robot to move well means coordinating the whole body, reacting in real time, and not falling apart when the environment changes. Behavior Foundation Models try to handle all that by training on large amounts of motion data, but it has not been clear what actually makes them scale. This paper works through the recipe.
The authors point to three pieces that need to work together: framing many control problems as tracking and reproducing whole-body motions in a global frame, balancing how much on-policy rollout data you collect against how varied the reference motions are, and an architecture they call the Humanoid Transformer. In simulation and on real hardware, their keypoint position error drops by over 10 percent in local mode and 82 percent in global mode against existing controllers.
Those figures come straight from the abstract, so read the paper for how the modes differ and how error was measured.
Humanoid control requires natural whole-body coordination, precise real-time responses to control signals, and robust generalization across diverse environmental contexts, making it a cornerstone for generalist embodied agents. Behavior Foundation Models (BFMs) have recently emerged as a promising solution to address these challenges by leveraging large-scale behavioral data to achieve superior expressiveness, versatility and generalization. However, despite growing interest…
Evaluating Open-Weight LLMs for Generating Structured Threat Information for Autonomous Vehicle Vulnerabilities
arXiv (cs.AI) · July 17, 2026When Does Muon Help Agentic Reinforcement Learning?
arXiv (cs.LG) · July 17, 2026Behaviour-Conditioned Neural Processes for Adaptive Residential Short-Term Load Forecasting
arXiv (cs.AI) · July 17, 2026An Exam for Active Observers