Ep#23 FALCON - Learning Force-Adaptive Humanoid Loco-Manipulation

In this talk

1. 1:43

   Why heavy contact challenges humanoids

   Han says many humanoid projects focus on lighter tasks, while pushing or pulling objects creates difficult forces across both arms and legs. The goal was to keep walking stable and arm motion accurate despite unknown forces from different directions. He found that a single whole-body reinforcement learning controller struggled to balance those goals and needed more reward tuning.

2. 6:04

   Two agents and a force curriculum

   FALCON trains separate upper-body and lower-body actor-critic policies together, with both using whole-body proprioception. During training, it samples forces within ranges calculated from the robot’s joint positions and torque limits. Han says the method trained in roughly 6,000 iterations, or two to three hours, and the same reward settings worked on Unitree G1 and Booster T1 robots.

3. 10:12

   Physical demonstrations and limits

   In demonstrations, the robots resisted pulling forces, opened a door, and pulled a cart weighing about 85 kilograms. Han reports peak stance resistance of about 60 newtons and pulling resistance near 40 newtons. He also describes hardware constraints: a 3.7 kilogram bottle exceeded a wrist’s stated three kilogram torque capacity, and the robot could hold it only briefly before its motors softened.

4. 12:56

   Autonomy and what remains difficult

   The autonomous tote-carrying pipeline uses FoundationPose to locate grasp points, planning for the arm and walking motions, and FALCON for control. Han says the controller was trained on flat ground; stairs and other uneven terrain remain future work. Carrying heavy or shifting loads can also cause vibration, which the team had not yet resolved.

5. 31:35

   When to decouple control

   Han says FALCON uses no reference character and transfers from simulation to the robots. He recommends splitting arm and leg learning for stable locomotion while the arms track poses, but says whole-body training fits tasks such as football-like motion tracking better. He suggests applying a carefully designed force curriculum when adding disturbances to motion-tracking systems.