It’s routine for four-legged robots with computer vision to navigate stairs, but getting a “blind” bipedal robot to do it is a whole other challenge. Now, researchers from Oregon State University have accomplished the feat with a bipedal robot called Cassie (from Agility Robotics) by training it in a simulator. 

Why would you want a blind robot to navigate stairs? As the researchers point out, robots can’t always rely completely on cameras or other sensors because of possible dim lighting, fog and other issues. So ideally, they’d also use “proprioception” (body awareness) to navigate unknown environments. 

The researchers used a technique called sim-to-real Reinforcement Learning (RL) to establish how the robot will walk. They noted that that “for biped locomotion, the training will involve many falls and crashes, especially early in training,” so a simulator allowed them to do that without breaking the robot. They taught the robot virtually to handle a number of situations, including stairs and flat ground. 

With simulated training done, the researchers took the robot around the university campus to tackle staircases and different types of terrain. It proved to be an apt pupil, handling curbs, logs and other uneven terrain that it had never seen before. On the stairs, the researchers did 10 trials ascending stairs and 10 descending, and it handled those with 80 percent and 100 percent efficiency, respectively. 

There were a few caveats in the first trials, as the robot had to run at a standard speed — it tended to fail if it came in too fast or too slow. It’s also highly dependent on a memory mechanism due to the challenge of navigating an unknown environment while blind. The researchers plan future tests to see if the efficiency improves with the addition of computer vision. All told though, “this work has demonstrated surprising capabilities for blind locomotion and leaves open the question of where the limits lie,” they wrote.