Have you ever skied down an immaculate white slope? Hard to see the bumps, right?
The same is true for the Yeti robot that needs to drive through polar regions that feature obstacles, slopes and different densities of snow. In such low-contrast terrain, vision won’t be able to detect challenging situations that might get the robot stuck. Instead, robots should rely on proprioceptive sensors, such as gyroscopes, accelerometers, motor current and wheel encoders to indirectly ‘feel’ the terrain below.
Using this idea, Trautmann et al. developed an algorithm that makes the robot learn to detect what it ‘feels’ like right before getting stuck (using a Support Vector Machine). The dangerous situations are then classified (using a Hidden Markov Model) and an escape behavior is implemented.
Polar terrain features that present a mobility challenge to the 73kg Yeti robot were determined during field deployments in Greenland and Antarctica. These challenging scenarios were reproduced in Hanover and used to train the robot. Results show that the robot is able to detect tricky situations with an error rate as low as 1.6% for a variety of obstacle geometries, approach angles to obstacles, robot speeds, and snow conditions. Furthermore, the robot is able to recognize the challenge type correctly in 100% of situations.