ForPC control lets a Unitree G1 humanoid push a 30 kg cart without force sensors

A team led by ZhenYu Zou has described ForPC in an arXiv paper, a whole-body control method that lets a humanoid robot push a wheeled cart while keeping its balance. Pushing a cart with both hands is hard because the forces that move the cart also unsettle the robot, and a heavier load changes how the two move together.

ForPC uses two stages of reinforcement learning: a walking policy first learns to keep balance under different hand loads and postures, then a second policy estimates forces from the robot's own sensors and adjusts arms, waist and legs. In simulation it cut cart speed error by up to 68.2% compared with walking control alone, and on a real Unitree G1, with no extra training, it moved a cart loaded with up to 30 kg at 0.5 m/s.