A reader pointed me to this article about robot toddler that learns to walk like a human child. This achievement is an excellent example of how artificial intelligence techniques can be used to reap efficiency gains that cannot be easily engineered by human hands.
The machines use what the researchers called a "passive-dynamic design" that closely mimics the way humans walk. Earlier robots required powerful machines to stroll, with each leg, knee and ankle requiring motorized assistance. The effort requires a lot of energy.
The passive dynamic design uses gravity, along with muscle-like springs and motors. The energy required is just a fraction of that needed by other walking robots, said Andy Ruina, a Cornell University researcher.
Ruina said the walking robots move like humans, falling and catching themselves as they move forward. This essentially is the same movement people use, a motion toddlers must master to walk.
"We let the machines take care of a lot of the motion," he said. In contrast, most walking robots, such as Asimo, developed by the Honda Motor Co. (HMC), require a motor to power every motion.
Could a control system for this kind of movement be designed by hand? Probably, but it would be incredibly hard, particularly if it were to be as adaptable as this learning model.
"It can learn to walk in 20 minutes," Tedrake said. "Once it learns to walk, then it adapts its gait to new terrain."
He said the sensors take measurements at the rate of 200 times a second and constantly send new instructions to the motors that control the tilt and motion. The sensors also direct actuators that control the tension on springs in the robot ankles. This helps the machine push forward with each stride.
"Every time it takes a step, it changes the parameters a little bit, based on its experience," Tedrake said. "It will walk on any surface and adjust the way it walks."
In effect, the robot changes its stride just as humans do when moving from sand to grass to pavement.
He said the machine even has learned to walk on a treadmill, making adjustments as the surface tilts or speeds up. The robot can start on its own and even walk backward.
I have a little experience with robotics, and these results are pretty impressive. Artificial intelligence is very well suited for developing control systems like these walkers, and there's more promise in this direction than towards the kind of AI you see in science fiction books and movies.