A paper presents ReACT, a method for keeping a humanoid robot's walking speed accurate when it carries an object in one hand. The uneven load pulls the robot off course, and small errors can add up to collisions.
The method uses the estimated force from the load to adjust the controller, and trains by comparing loaded and unloaded runs with the same speed commands. In simulation on a Unitree G1, speed tracking improved and sideways bias fell under different loads.
In real-world tests with a load in one hand and a fixed navigation policy that uses no map, the robot reached its goal without collisions in all 40 trials.
Background
ReACT
Humanoid control method
A training framework that keeps a humanoid's walking speed accurate when it carries a load in one hand.
