FORGE: Feasible Online Reduced-Order Reference Generation and Whole-Body Execution for Agile Humanoid Locomotion
SUMMARY
Adrian B. Ghansah, Yujin An, Jeeseop Kim and Soon-Jo Chung have posted a paper on arXiv describing FORGE, a two-stage learning framework for agile humanoid locomotion. It links online reference generation from a reduced-order model (ROM) with a whole-body tracking controller. The authors submitted it on 27 October 2026, and it has been accepted to the 2026 IEEE-RAS 25th International Conference on Humanoid Robots (Humanoids).
In the first stage, the authors train a reference generator that takes commands as input and produces whole-body references. It is built on single rigid body dynamics combined with the full kinematics of the humanoid. In the second stage, the reference generator is frozen and rolled out online, and the references it produces are used to train a whole-body tracking policy on the full-order dynamics model.
The authors say that learning a kino-dynamically consistent reference generator in the simplified ROM space lets them produce references across a broad command space and across several locomotion strategies, within a single framework.
They demonstrate the method on a Unitree G1 humanoid robot. Reported results include longitudinal velocities from -2.0 m/s to 2.8 m/s, lateral velocities up to +-1.0 m/s and angular velocities up to +-3.0 rad/s, as well as one-legged hopping and continuous transitions between commands. According to the abstract, one reference policy and one whole-body tracking policy cover all of these behaviours.
Drafted with AI assistance from the source and reviewed by WBH. Follow the source link for the full text.
WHY IT MATTERS
EDITORIALFORGE couples a learned reduced-order reference generator with a whole-body tracking policy. According to the abstract, this single framework covers a broad command space, several locomotion strategies including one-legged hopping, and continuous transitions between commands.