Combined machine and human learning facilitates fast robot turns on steep granular slopes.

Planetary exploration often requires reaching challenging regions, such as craters and steep slopes. Granular slopes are prone to flow under small disturbances, making rover traversal challenging. To discover principles by which a new class of four legged and wheeled rover (able to lift, sweep, and spin its wheels, see Srivastava et al, 2020) can effectively turn in place on granular slopes without toppling over or sliding down the slope, we studied in the laboratory a 30 cm long robophysical rover model on a granular medium of poppy seeds. The medium was contained in a tiltable fluidized bed which could achieve slope angles of 30 degrees. Because of the many control parameters in the robot, systematic sweeps of parameter space proved challenging. Therefore, we used the machine learning (ML) technique of Bayesian Optimization to target successful turning gaits in the gait search space and found strategies such as differential wheel spinning and pivoting around a single sweeping wheel. This resulted in a gait which allowed the robot to turn 90 degrees in approximately 2.5 minutes without toppling or sliding. By studying how this ML gait generated efficacious interactions of the robot and the material, we used human learning (HL) to further modify the turning gait. This resulted in a new gait which enabled the rover to turn 90 degrees at 4.5 seconds with minimal slip. The combined ML and HL approach generated fast turning by creating anisotropic drag on the sweeping wheel during different phases of limb lifting.