Watch the ‘blind’ robot managed to navigate the stairs

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

Why do you want blind robots to navigate the stairs? When the researchers show, robots do not always depend entirely on other cameras or sensors because of the possibility of dim lighting, fog and other problems. So ideally, they will also use “proprioception” (body awareness) to navigate an unknown environment.

The researchers used a technique called Sim-to-Real Reverforcement Learning (RL) to determine how robots would run. They noted that “for biped drivers, training will involve many falls and crashes, especially the initial training,” so the simulator allows them to do it without damaging the robot. They taught robots virtually to handle a number of situations, including stairs and flat land.

With simulation training done, the researchers took robots around the university campus to handle stairs and various types of Medan. It proved to be an APT student, handle the uneven sidewalks, logs and fields so they have never seen before. On the stairs, the researchers conducted 10 trials climbing stairs and 10 down, and each handle them with an efficiency of 80 percent and 100 percent.

There are several warnings in the first trial, because the robot must run at standard speed – it tends to fail if it’s too fast or too slow. It is also very dependent on the memory mechanism because the challenge of navigating an unknown environment during blind. The researchers plan a future test to see if their efficiency improved with the addition of computer vision. All told me, “This work has shown a surprising ability for blind drive and leaves opening questions about where the boundaries lie,” they wrote.