Researchers Showcase Self-Learning Quadrupedal Robot for Which Ladders Pose No Challenge
The Tokyo Metropolitan University and Okayama University's compact quadruped is very good at climbing one specific ladder, at least.
Researchers at Tokyo Metropolitan University and Okayama University have demonstrated prototype quadruped robots with a feature rarely seen: the ability to climb ladders, using 3D sensor input and machine learning algorithms running in simulation to figure out the best approach.
Presented at the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2019) conference earlier this month, and reported on in IEEE Spectrum, Azhar Aulia Saputra, Yuichiro Toda, Naoyuki Takesue, and Naoyuki Kubota's design is a quadruped with a difference: its legs feature five degrees of freedom each, providing it with the ability to switch from traditional quadrupedal motion to a gripper system which allows it to climb ladders — with one or two caveats.
The robot works through a combination of live data fed via dual laser rangefinder sensors, a depth-sensing camera system, and an inertial measurement unit, along with touch and pressure sensors located in the grippers on its feet, and simulation training which occurs in simulation. The latter, the researchers admit, needs work. The training is specific to the mock-up ladder used for the physical testing, meaning that if any variable of the ladder changes — width, height, distance between rungs, even rung thickness or shape — the robot would need to be re-trained.
The team's paper, A Novel Capability of Quadruped Robot Moving Through Vertical Ladder Without Handrail Support, has not yet been released publicly; more information, however, is available on IEEE Spectrum.