'Blind' robot learns to navigate a flight of stairs for the first time by feeling its way

 

  • Cassie had 80 percent success rate in climbing stars at Oregon State University
  • The bipedal robot uses body awareness, to proprioception, to navigate steps
  • That can be helpful for delivery robots under poor lighting conditions
  • Cassies previously walked through fire and rode a Seque scooter Engineers in the US have devised a robot that can easily climb staircases in the dark.

    'Cassie' ascended the steps at Oregon State University with 80 percent proficiency, all without eyes or other sensors.

    The bipedal robot was trained to use 'proprioception,' or body awareness—to navigate uneven surfaces.

    Researchers say that's important if fog, dim lighting, or other factors limit a robot's visual acumen.Scroll down for video

    'Cassie,' a bipedal robot, ascended and descended he steps at Oregon State University without computer vision or other sensors

    'Cassie,' a bipedal robot, ascended and descended he steps at Oregon State University without computer vision or other sensors

    Cassie was developed by the Dynamic Robotics Lab at Oregon State University in 2017.

    It has no 'head' but its hips have three degrees of freedom, allowing it to move its legs forward and backward, side-to-side, and also rotate them at the same time,' according to the Institute of Electrical and Electronics EngineersIn addition, Cassie's powered ankles allow it to stand in place without constantly having to move its feet and shift its weight.

    It's classified as a 'dynamic walker' with a smoother, more human-like gait than your typical thudding automaton to make it more adept at traversing complex terrain—well, complex from a robot's point of view.Cassie is classified as a 'dynamic walker' with a smoother, more human-like gait than your typical thudding automaton to make it more adept at traversing varied terrain

    Cassie is classified as a 'dynamic walker' with a smoother, more human-like gait than your typical thudding automaton to make it more adept at traversing varied terrain

    Cassie already withstood other extreme gauntlets, walking through fire and successfully riding a Segue.

    Researchers at Oregon State were interested in training Cassie to climb stairs 'blind,' that is with no computer vision or other sensors.

    Robots have long been able to conquer staircases using cameras and computer vision but certain conditions, like dim lighting, aren't always ideal for visual input.

    They wanted Cassie to go up and down the steps using only its 'proprioception,' or body awareness—the same way you or I might creep down to the basement at night.

    Before Cassie could be put to the test on an actual flight of stairs, though, engineers trained it virtually,  the researchers explained in a paper posted on the open-access platform ARXIV, with a technique called 'sim-to-real reinforcement learning.'

    In 10 trials, Cassies had 80 percent success rate ascending the stairs and 100 percent success descending

    In 10 trials, Cassies had 80 percent success rate ascending the stairs and 100 percent success descending 

    Cassie was trained to climb steps and maneuver other surfaces with 'sim-to-real reinforcement learning, which uses a simulator to show it how to handle a number of situations, thereby avoiding 'many falls and crashes, especially early in training'

    Cassie was trained to climb steps and maneuver other surfaces with 'sim-to-real reinforcement learning, which uses a simulator to show it how to handle a number of situations, thereby avoiding 'many falls and crashes, especially early in training'

    Using a simulator, they taught Cassie how to handle a number of situations, including stairs and flat ground, thereby avoiding 'many falls and crashes, especially early in training,' the team said.

    Eventually it was time for a real-world test: The researchers took Cassie OSU campus and watched as it tackled curbs, staircases and other uneven surfaces it had never encountered before.

    In 10 trials, Cassies had 80 percent success rate ascending the stairs and 100 percent success descending.

    Cassie did have a few limitations, including needing to stay at a constant speed. It tended to mess up if it came at the stairs too fast or too slow

    Cassie did have a few limitations, including needing to stay at a constant speed. It tended to mess up if it came at the stairs too fast or too slow

    'To our knowledge, 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,' the researchers said.

    Cassie did have a few limitations, including needing to stay at a constant speed. (it tended to mess up if it came at the stairs too fast or too slow, engadget reported.)

    'This work has demonstrated surprising capabilities for blind locomotion and leaves open the question of where the limits lie,' the team wrote.

    Three months after being turned on  Cassie was able to learn to walk on a variety of terrain without falling over —a track record most human can't match at that age

    Three months after being turned on  Cassie was able to learn to walk on a variety of terrain without falling over —a track record most human can't match at that age

    Ultimately, Agility Robotics sees Cassies as a combination deliveryman and pet dog, scampering up the steps to deliver packages to your front door, even if it's foggy or dark.

    They've had remarkable success: After just three months, Cassie was able to walk without falling over on a variety of terrain, a track record most human bipeds couldn't match at that age.

    As of October Agility has raised $29 million in funding, with Sony and TDK among its backers.

    Cassie no 'head' but its hips have three degrees of freedom, allowing it to move its legs forward and backward, side-to-side, and also rotate them both at the same time

    Cassie no 'head' but its hips have three degrees of freedom, allowing it to move its legs forward and backward, side-to-side, and also rotate them both at the same time

No comments:

Powered by Blogger.