Jumping to the Moon: A strengthening learning policy guides a quadruped robot in lunar gravity

A single control policy, trained in several stages and limited to embedded sensors, allows precise point-to-point jumps in simulation, and a first vertical jump was made on a real robot.

Legged robots could explore lunar terrain that is too rough for wheeled rovers, and the low gravity gives them an additional opportunity to jump over obstacles or crevices that are inaccessible on foot. Levin Müller, Oliver Fischer, Marco Trentini, Yusuke Tanaka and Marco Hutter present a multi-stage learning method that results in a single, end-to-end control policy capable of making a jump to a point commanded in lunar gravity. The architecture remains simple and relies only on embedded sensors compatible with a real mission, making a deployment on the Moon plausible according to the authors.

In simulation, the robot lands with an error less than the size of its own body. The team also studied the torque and power requirements of these jumps and the degradation of performance in extreme situations, such as a control loop freeze of up to a second in full flight. On the physical robot, a device compensating part of the weight to reproduce the lunar gravity made it possible to pass a vertical jump; the measured torques and powers are consistent with the simulation and remain within the limits of the material.