MIT’s Computer Science and Artificial Intelligence Lab (CSAIL) has a new robotic trick up its sleeve. teaching robots to teach other robots. Researchers at the school have figured out a new system that allows non-experts (i.e., people who aren’t roboticists) to teach robots how to do simple (by human standards) behaviors, including opening doors and moving things around. The real twist, though, is that the robots can then also teach those same tasks to other robots.
Yes, robots teaching robots. The so-called C-LEARN program developed by the researchers at CSAIL combines prior knowledge with simple instructions around a specific task, provided by a human bot teacher. The robot’s skills are then transferable, one to another, meaning theoretically you could have a single, non-specialist instructor provide the basic instructions that can end up educating an entire army of robots.
Those robots don’t have to even be exactly the same as the teacher model; during research, the Optimus model you see in the video above was able to pass on learned actions to the larger Atlas model you see in the video below.
The key to the system’s success is that it combines demonstration with an existing knowledge base provided prior to training. These result in high degrees of accuracy in accomplishing the demonstrated task, which are then translated to subsequent teaching to other robots.
The C-LEARN system is far from perfect, however; teaching a robot how to pick up a box can take as long as half-an-hour, and complex tasks, including collision avoidance aren’t yet teachable via the system. Still, with enough work, it could be an effective cargo loader or even a simple maintenance workforce.