JUST IN: A team from the University of Bristol has presented a robotics technique that aims to shift some of the control effort from the software to the robot's own body.

In simulations with a Unitree Go2 quadruped, the system managed to offload up to 87% of the mechanical power to passive elastic joints on flat terrain, with net savings of up to 86% compared to a baseline without this physical support.
The method, called Physical Imitation Learning, divides the control policy between motors and passive elastic joints.