The "motion challenges" of humanoid robots may now have a new solution.

Horizon has just officially announced: its robot laboratory has developed the HoloMotion-1 model, which is now open-sourced. This is a "cerebellum" large model with 400 million parameters, focusing on full-body control of humanoid robots. As the name suggests, it does not handle "thinking," but rather focuses on "moving stably, accurately, and human-like."

This open-sourced HoloMotion-1 is actually the "first step" in Horizon's HoloMotion technology roadmap. Its core mission is very focused: to overcome the ability to "imitate any pose."

What does that mean? It means that robots can "learn" movements from diverse data sources:

🎥 A video of a human → can decompose movement logic

🕺 MoCap motion capture data → can restore joint trajectories

🎮 Remote operation real-time commands → can synchronously reproduce complex postures

In other words, in the future, training robots to perform full-body coordinated actions such as rolling, grasping, and walking will no longer rely entirely on engineers writing code line by line, but rather let the model "understand" movement patterns from the data itself.