We tend to think of four-legged robots as robotic versions of dogs. And, to be honest, it is present right in the word “four-legged.” But if we can just get around Latin, then there’s absolutely no reason why quadrupedal robots should limit themselves to using all four limbs as legs all the time. And in fact, most other tetrapods are just as versatile: tetrapods often use their forelimbs to interact with the outside world for purposes other than locomotion.

Roboticists from CMU and UC Berkeley are training robotic dogs to use their legs for manipulation, not just locomotion, demonstrating skills including climbing walls, pushing buttons and even kicking a soccer ball.


Teaching a robot to move and manipulate the same limbs at the same time can be tricky with reinforcement learning methods because you can get stuck in local minima trying to optimize skills that are very different and (I guess) sometimes in opposition to each other. So the researchers split the learning into separate manipulation and locomotion policies and trained each of them in a simulation, although that meant an extra step of bringing those separate skills together in the real world to perform useful tasks.

Successfully completing a combined locomotion and manipulation task requires one high-quality expert demonstration. The robot remembers what commands the human gave during the demonstration and then builds a behavior tree that it can follow by breaking down the tasks into a set of interrelated movement and manipulation subtasks that it can perform in order. It also increases the reliability of the system, because if the robot fails on any subtask, it can “rewind” its way back through the behavior tree until it returns to the success point, and then start from there.

This particular robot (Unitree Go1 with Intel RealSense for perception) manages to balance itself against a wall, press a wheelchair access button nearly a meter high, and then exit through an open door, which is pretty impressive. More broadly, this is a useful step towards helping non-humanoid robots operate in a human-optimized environment, which may be more important than it sounds. It is certainly possible to change our environment to be more robot-friendly, and we see this in places like hospitals (and some hotels) where robots can directly control elevators. This makes it much easier for the robots to move around, but in some cases it’s annoying enough that it’s more practical (if not necessarily easier) to just build a button pushing robot instead. It might be argued that the best middle ground here is to simply build a public infrastructure in the first place, making sure that neither robots nor humans have to rely on a particular manipulation technique to control anything. But until we do, such skills will be critical for useful legged robots.

Legs as a manipulator: quadrupedal dexterity goes beyond locomotioncreated by Shuksin Cheng, Ashish Kumar and Deepak Patak of Carnegie Mellon University and UC Berkeley, will be unveiled next month at ICRA 2023 in London.

From articles on your site

Related articles online