Swiss Mile’s ANYmal robot is a remarkable beast, capable of getting around as a wheeled quadruped, or standing up on its hind legs and using its front wheels as hands. Now, it’s learning to do useful tasks – in about the funniest way possible.
The architecture here is pretty phenomenal in its own right; ANYmal is a great name for this thing. It looks like a robot dog in quadruped mode – albeit one with rather fiddly looking legs, as well as the odd inclusion of wheels. But that makes for super-efficient locomotion across a range of even and uneven surfaces, including stairs. Why walk when you can roll – and indeed, why waste energy on bipedal balancing when you can get around on four legs much more easily?
But, as the humanoid-focused companies point out, a lot of the tasks we want robots to take over are currently designed for the human form. And when it needs to act more like a human, ANYmal simply squats back and stands up, using its motorized wheels to self-balance. Here, look at this video from a year ago:
Advanced Skills through Multiple Adversarial Motion Priors in Reinforcement Learning
Standing up leaves its front legs free to act as arms. And while it would be totally possible to have a set of hands tucked away that could fold out for dextrous manipulation in humanoid mode, the team at Swiss Mile – an ETH Zurich spinout company – has decided to see just how much this bot can get done using nothing but its powered wheels as hands.
One of the earliest such tasks, as shown in the video above, was pressing elevator buttons – fairly remarkable in itself when you consider the precision that requires as opposed to the rather blunt instrument of an 8-10-inch (20-25-cm) diameter wheel.
While the above video had the ANYmal approaching tasks from reinforcement learning and imitation learning perspectives, and attempting to reverse certain motions in order to achieve the opposite task, the team has more recently been experimenting with a new technique it’s calling “curiosity-driven learning.”
Swiss Mile
In the new approach, the robot is given a task, rewarded only when it completes the task, and encouraged to explore and play with goal-related items in its environment, effectively being told to just get in there and figure things out for itself.
So nobody needs to go in there and do the tedious work of designing complex reward schemes to guide the robot toward its goal. And nobody needs to sit there and demonstrate the task a hundred times over so the robot can watch and learn. You just need to set some key variables, point out objects that could be significant to the task, and give the robot its final goal.
Swiss Mile
In testing, this technique is already delivering impressive results – the ANYmal learned to open doors well enough to nail the task 15 times in a row, for example, and also figured out how to pick up a box and put it in a bin – which is going to be one of the key early jobs needed from general-purpose robots as they start entering the workforce.
But the fun here is in how the ANYmal gets the job done – given the simple goal of “get box into bin,” it’s learned to pick them up and fling them at the bin in a way that would make airport baggage handlers proud. There’s something wonderfully… relatable, perhaps, about the way it chucks those boxes. Enjoy the video below.
Curiosity-Driven Learning of Joint Locomotion and Manipulation Tasks
At this stage, ANYmal is more of a research project than anything else. But Swiss Mile could move to commercialize it, and this bot’s remarkable transforming capability could be a real game changer, combining the advantages of a humanoid form with the compact, efficient locomotion of a wheeled quadruped. Very neat work!
Source: Swiss Mile via IEEE Spectrum
