The robots, especially quadrupeds, equipped with cameras and sensors already demonstrated their ability to climb stairs. Spot, from Boston Dynamics, is one of the paradigms in this field. But getting a “blind” bipedal robot to do it is a whole other challenge.
Now, a team of researchers from Oregon State University has accomplished the feat with a bipedal robot called Cassie from Agility Robotics. The team trained the blind bipedal robot to navigate stairs – without any perception sensors such as LiDAR or cameras – in a simulator.
Researchers say robots can’t always rely completely on cameras or other sensors. With such possible conditions as dim lighting or fog, movement can be challenging. To solve this issue, Agility Robotics’ bipedal robot ‘Cassie’ is trained to navigate an unknown environment through ‘proprioception,’ or body awareness.
The team used the technique called sim-to-real Reinforcement Learning (RL), which virtually established how Cassie will walk. For such blind bipedal locomotion, the training will involve many falls and crashes, especially early in training. To avoid this, the simulator allows the robot to train without damaging itself.
As we can see in the video below, they virtually taught the robot how to handle a number of situations, including stairs and flat terrain, then brought it into the real world. Cassie demonstrated the ability not only to climb stairs but also to avoid logs and move on uneven terrain that it had never walked before. Regarding the ascent and descent of steps, it achieved an efficiency of around 90% in the tests.
After successful physical tests, the team notes, ‘this is the first controller for a bipedal, human-scale robot capable of reliably traversing a variety of real-world stairs and other stair-like disturbances using only proprioception.’
