Robot learning method improves transport of unsecured stacked boxes

Researchers report in an arXiv preprint that their PAMORT learning method improved a quadrupedal robot's ability to carry unsecured stacked boxes. Across eight real-world tasks on a Unitree Go2, mean success rates were 0.850 for PAMORT and 0.675 for the baseline. The method adjusts the balance between walking performance and payload stability using the robot's own sensing of its body, without dedicated payload sensors or active carrier mechanisms. Real-world tests transferred the trained policy to slopes and steps at or beyond the training difficulty. In simulation, it also handled an unseen three-box stack despite training only with two boxes.