Researchers from ETH Zurich have developed a learning-based controller that enables a quadrupedal robot with an arm to push and reorient unknown objects with high accuracy.

Using constrained reinforcement learning, the system adapts to different object properties-mass, material, size, and shape-while achieving a 91.35% success rate in simulation and over 80% on hardware.

The robot dynamically adjusts its pushing strategy, ensuring stable and contact-rich interactions without prior object knowledge