One day in the near future you could be lying incapacitated in the middle of an earthquake-ravaged landscape where wheeled vehicles can’t reach you. Or in the very distant future, after engineers have solved thousands of problems associated with saving interplanetary astronauts from radiation, your great-granddaughter could be lying incapacitated in the middle of a marsquake-ravaged landscape where wheeled vehicles can’t reach her.
So, who’s going to save you (or her)?
If it’s up to an international group of researchers led by Japan’s Tohoku University and Thailand’s Vidyasirimedhi Institute of Science and Technology (VISTEC), it might just be a massive insectobot – one that learned to walk over different surfaces through AI-powered biophilic copying of stick insect locomotion.
“It’s remarkable that a few steps from a single stick insect,” says Dai Owaki, Associate Professor at Tohoku University, “were enough to find a principle that works on a machine five times its size.”
In their Bioinspiration & Biomimetics paper, Owaki’s team explains that while insects have tiny brains and nerve networks compared to humans, they are still capable of remarkably complex movement and dynamic adaptation. But until now, entomologists and others have been unable to identify and computationally model the exact principles of insect leg coordination.
The result has been imperfect biophilic (or insectomimetic) coordination rules, parameters, and reward functions that hobble flexibility, and are so specific that they prevent transfer from one body-type of robot to another.
From Insect Behavior to Transferable Robot Locomotion
As Owaki explains, his team departed from traditional approaches for teaching robots how to move. “We never told the robot how to walk. We asked what the insect was trying to achieve, and let the robot chase the same thing entirely on its own.”
Because he and his colleagues were working to create a teaching model that didn’t rely on supplying the AI with explicit instructions on how to walk, they used adversarial inverse reinforcement learning – that is, an imitative approach that identifies the “reward” of safe foot placement while adapting to changing conditions and identifying new rewards.
The outcome was so successful that the hexapedal training robot needed only brief exposure to a walking stick insect navigating flat areas to “learn” how to walk over various terrains (and in future, other highly coordinated creatures could be the exemplars). Remarkably, it learned how to do so in under an hour.
Furthermore, the reward network was so robust that it could serve robots with a range of body types without requiring finicky customization. With added memory, say the researchers, robots using this learning method may one day be able to aid in disaster zones where smooth, even roads may not be accessible or may not even exist. And even if such robots lose limbs in the chaos, their learning system may be able to help them walk anyway.
If you’re fascinated by the emerging world of walking robots, you’re living in the best time (yet) of human history, and can enjoy Russian ostrich-bots, the rigid-legged bipedal Mugatu robot that looks absolutely nothing like the Mugatu, a massive hexapedal kaiju-mech, a robot dragon, and, of course, kung fu robots designed by people who never bothered to listen to Sarah Connor.
Source: Tohoku University
