Mirroring Motion
At the Advanced Telecommunications Research Institute in Kyoto, Japan, a million-dollar humanoid robot is learning to play air hockey. Using stereo video cameras, “DB” watches as a researcher strikes the puck with his paddle. Then using its hydraulically powered arm, the robot imitates the motion. After a few false starts, DB is able to hit the puck, and its movements are surprisingly graceful. This sort of “imitation learning” is yielding smarter, more adaptive robots for physical therapy, search-and-rescue missions, and space applications.
Imitation learning combines artificial-intelligence software with cutting-edge neuroscience. To learn arm movements for air hockey-or even for hitting a tennis ball-the robot uses machine vision algorithms to determine the position and velocity of a person’s limbs and maps this information to its hydraulic joints. Comparing its own movement with the original, the robot makes adjustments in real time. This is a more efficient way to make a robot perform human tasks than brute-force programming or trial and error, says computer scientist Stefan Schaal of the University of Southern California. Schaal is teaching DB new tricks in collaboration with the Japanese research group, Carnegie Mellon University, and Sarcos, a robotics company in Salt Lake City.
Until recently, however, “there was a large component of Simon Says with these robots. They didn’t understand what they were doing,” says Carnegie Mellon robotics expert Chris Atkeson. Changing one of their familiar tasks even slightly, in other words, would befuddle the machines. To add flexibility, researchers are now teaching robots to divide movements they’ve seen before into pieces that serve intermediate goals. Robots can learn to splice these building blocks together to adapt their behaviors-reaching, balancing, and even walking-to new situations. Within a few years, says Schaal, such resourceful robots could perform hazardous rescue missions, analyze stroke patients’ movements to help them refine motor skills, and replace space-walking astronauts on repair missions.
Keep Reading
Most Popular
Large language models can do jaw-dropping things. But nobody knows exactly why.
And that's a problem. Figuring it out is one of the biggest scientific puzzles of our time and a crucial step towards controlling more powerful future models.
The problem with plug-in hybrids? Their drivers.
Plug-in hybrids are often sold as a transition to EVs, but new data from Europe shows we’re still underestimating the emissions they produce.
How scientists traced a mysterious covid case back to six toilets
When wastewater surveillance turns into a hunt for a single infected individual, the ethics get tricky.
Google DeepMind’s new generative model makes Super Mario–like games from scratch
Genie learns how to control games by watching hours and hours of video. It could help train next-gen robots too.
Stay connected
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.