FALCON

A 2025 CMU humanoid training framework that lets a robot push, pull, and carry heavy loads steadily while walking.

FALCON is a whole-body control method for humanoid robots released in May 2025 by CMU's LeCAR Lab (Guanya Shi's group) together with Field AI and others, accepted as an oral presentation at L4DC 2026. When a humanoid walks and works at the same time (loco-manipulation), a large force on its hands — from pulling a cart, opening a door, or carrying something heavy — can easily throw the lower body off balance, while the upper body also struggles to reach its target accurately. FALCON uses dual-agent reinforcement learning: a lower-body policy keeps walking stable under force disturbances, and an upper-body policy sends the hand to its target while implicitly compensating for the external force, both trained jointly and sharing proprioception (joint angles, velocities, and other self-state); this is paired with a 3D force curriculum that gradually increases the force applied to the hands during training, while staying under each joint's torque limit. Upper-body tracking accuracy is about twice that of baselines, and the same pipeline, with no reward changes, works on both the Unitree G1 and the Booster Robotics T1.

ExampleIn real-robot experiments, the humanoid walked while pulling a cart under 0–100 N of force, opened a door with both arms against 0–40 N of resistance, and could carry a load while walking, squatting, and turning.

Also called

Learning Force-Adaptive Humanoid Loco-Manipulation

Related

Loco-manipulation · Whole-Body Control · Curriculum Learning · Unitree G1 · Booster Robotics T1 · Field AI

Sources

FALCON: Learning Force-Adaptive Humanoid Loco-Manipulation (arXiv 2505.06776)
FALCON 项目主页（CMU LeCAR Lab） (Chinese)

As of

2025-11