Humanoid Robots Learn to Push Carts Without Wobbling

Humanoid robots may soon take over the heavy, repetitive job of pushing carts in factories, warehouses, and hospitals. But teaching a two-legged robot to push a wheeled cart is tricky: when the robot grabs the cart, the forces from pushing can throw it off balance, and if the load changes, the robot has to adapt quickly. A new system called ForPC solves this by using a two-stage AI training process that helps the robot stay stable while pushing.

ForPC first learns basic balancing skills by practicing with different hand loads and body postures. Then, it uses sensors on the robot to estimate the forces involved and adjusts the robot's arms, waist, and legs to keep the cart moving smoothly. Unlike some systems, it doesn't need extra force sensors or human demonstrations. The AI is trained entirely in a computer simulation, but it works on a real robot without additional training.

In tests, ForPC reduced cart velocity errors by up to 68% and cross-track errors by 74% compared to a baseline that only controlled walking. The robot, a Unitree G1, successfully pushed a cart carrying up to 30 kg at 0.5 meters per second—about a slow walking pace—without any payload-specific adjustments. This means the robot can handle different weights without reprogramming.

The breakthrough suggests that humanoid robots could soon replace humans in strenuous material handling tasks, making workplaces safer and more efficient. While the tests are preliminary, the ability to transfer from simulation to reality without fine-tuning is a promising step toward practical deployment.