If you think about it, the fact that we often call quadrupedal machines “robot dogs” is pretty generous. Sure, they have four legs and can walk or crouch, but real dogs are miles ahead of their robot counterparts in maneuverability and interaction.
A trio of roboticists at Carnegie Mellon University and the University of California at Berkeley have developed a quadrupedal robot with a broader range of movement than your average robot dog. The robot combines agile locomotion with limb manipulation using its front legs as arms. The result is a robot dog capable of propping itself against a wall, hitting elevator and accessibility buttons, and even “playing” with a ball.
Their robot is a Unitree Go1 equipped with Intel RealSense, which enables 3D imaging. The team will present the robot at this year’s IEEE International Conference on Robotics and Automation (ICRA), but a video shared by one of the roboticists shows off its capabilities. In its first demo, the robot walks up to a wall and props itself up using its front limbs (much like an overexcited dog might when someone rings the doorbell). Keeping itself upright using one “arm,” the robot uses its other arm to press a button that opens the door to its left. It then gets back down on all fours and backs away before walking over the threshold.
Another demo shows the robot using its body weight to press larger buttons, like the long door-control buttons placed for wheelchair users. In its cutest (or most terrifying, depending on how you see it) demo, the robot crouches over a small soccer ball and kicks it. Despite IEEE’s assertion that we shouldn’t always think of quadrupedal robots as “dogs,” this is perhaps the most dog-like any function-forward quadruped has been in a while.
A significant hurdle to creating versatile robot dogs is local minima, in which mobile robots get “stuck” trying to navigate an unfamiliar obstacle or perform two or more motions simultaneously. In their paper, the roboticists write that they divided the quadruped’s training into distinct manipulation and locomotion policies. This allowed them to use their 18-layer neural network to train the robot in two parts, helping it develop behavior trees that break up commands into sequential subtasks. If the robot fails at a subtask or gets stuck, it can simply “rewind” until it returns to a point of success.
“Robotic quadrupeds are still far behind their biological counterparts,” the trio says. “We take a step towards bridging this gap by training quadruped robots not only to walk but also to use the front legs to climb walls, press buttons, and perform object interaction in the real world.”
