Imagine you have lots of different robots. Some have four legs (quadrupeds), some have two (humanoids) and some are made like bugs (with 6 legs). While it's easy to understand that your quadrupeds would move differently from your humanoids and so on, in practice, even the same kind of robot would move differently based on the build and various factors. So the same type of robot needs to be separately trained. That is time-consuming and costly. If only there were a way, one single policy would train all these programmable robots to make robot locomotion simple and accessible.
In today's article, we'll take a look at the exceptional research done by a team of AI and robotics university researchers from Germany and Poland to solve the training problem by creating URMA, the Unified Robot Morphology Architecture, to close this gap with artificial intelligence, reinforcement learning, and multi-embodiment learning. You can access Nico's published paper in 2024: "One Policy to Run Them All: an End-to-end Learning Approach to Multi-Embodiment Locomotion" (Nico Bohlinger, Grzegorz Czechmanowski, Maciej Krupka, Piotr Kicki, Krzysztof Walas, Jan Peters, and Davide Tateo)
Back in the day, owning a robot was a luxury, but today, robots are now accessible wih affordable prices. Take, for instance, Bittle smart robot dog (which was also used as part of this research), which can be adopted for less than 300$. As robots move from labs to people's homes, there are more platforms and more complexity on the backend. To tackle these issues, this research paper attempts to create one smart brain (also known as a policy) that can train all kinds of robots to learn to walk, no matter how many legs they have or how they’re built. As you read through the research paper, you will see that the researchers trained this super brain first by using a computer simulation. They added lots of different robot shapes, taught them all at once, and then tried to see if the smart brain could handle new robots it had never seen before. And guess what? It worked pretty well! The next obvious step was testing it on actual robots. Here is a video that shows the progression:
If you'd like to watch a more in-depth explanation of the research that also dives into the technical aspects of the research, you can watch this full video from Nico: "One Policy to run them all".
Here is a snippet of Nico explaining the policy:
So what does this mean for the future of robotics development? As of today, when someone develops a robot, they have to start teaching their robot how to move from scratch. As mentioned earlier, this is slow and expensive. But through this research, we can teach many robots to move through the same shared system. This will save time, make robots adaptable, and reduce training costs as well! All this translates to smarter, smoother, and faster deployment of robots across various industries like medical, space, and civil infrastructure!
Note that the policy has done all the robot learning on even flat terrains. In the future, the research team plans to tackle tricky terrains like steps, grass, slopes, parkour, and also extend this training to humanoid robots! We're elated that Petoi open source robot dog Bittle could contribute to this research, and we wish the research team all the very best for the future!
If you're looking to get into AI robotics, check the numerous projects with Petoi robots:
