By Germain Chastel and Martial Chastel
The evolution of the human brain to what it is today took millions upon millions of years. The evolution of artificial intelligence has taken less than 100 years. It’s hard not to compare the two — after all, the goal of artificial intelligence is to reach general intelligence, or, the point at which a machine’s brain can successfully perform any intellectual task that a human brain can. It’s no wonder, then, that some of the biggest AI developers in the world have begun teaching machines through a process that is not so different from how nature taught humans.
NewtonX interviews with a former lead engineer at Boston Dynamics, an executive with Nvidia, and a former analyst with Google’s DeepMind revealed that the biggest robotics and AI innovators are using simulation environments coupled with Artificial Neural Networks (ANNs) to teach machines to evolve through trial and error in much the same way that all natural organisms have evolved. There’s just one difference: what took us humans centuries upon centuries to learn can be taught in a simulation environment in a matter of weeks.
The Pre-simulation Problem: It’s a big world for a little robot
One of the reasons that game environments are so appealing to AI developers is that they exist in an enclosed environment in which only a finite number of actions can occur. This means that development of an AI “player” is easier than, say, the development of a smart speaker or a physical robot. The number of possible things that a human could say to Siri is unlimited; the number of possible moves a Go player could make is limited. That’s why AlphaGo was able to beat a human, but Amazon’s Alexa still struggles with language disambiguation and understanding different idioms.
Because of these issues, most forms of AI today don’t move through the world in the same way humans do: they need limits in terms of space and interactions. Those who have attempted to broaden the scope of physical AI-powered robots have for the most either disappointed or failed (such as the comedic incident in which a DC security robot, Knightscope K5, drowned itself in a fountain).
As a senior engineer at Boston Dynamics put it, “The ground is a massive problem for a little robot. There are just too many obstacles moving around — forget cars and pedestrians; the problem could be as small as a fence, a ditch, rough terrain, or stairs that are just a little steeper than normal, and can completely derail a robot.”
But recently, according to NewtonX experts, simulation environments have provided a potential answer to the conundrum of teaching robots how to interact — intellectually and physically — with the surrounding world.
Simulation environments teach robots physical boundaries
Simulations provide a virtual model that can emulate real-world processes. In robotics, simulations can also create a virtual model of a robot, including the robot’s design and programming code.
One of the most commercially successful examples of a robot that was likely developed this way is the Roomba and other robotic vacuum cleaners. Developing them is relatively simple (relative being the operative word — iRobot spent over a decade researching and developing) because most houses have similar features: rugs, tile, concrete, or hardwood floors; rectangular or circular furniture, doors, and a few discarded items on the floor (clothes, toys, etc.). In a simulation environment, the robot can learn through positive and negative reinforcement how to interact with the world around it — in much the same way that children learn, and on a larger scale, in much the same way that humanity as a species learned.
Simulation environments allow this evolution to occur at exponential rates. Instead of building an entire robot, observing its failures, iterating, and doing it all over again, simulations allow this entire process to occur without building a physical robot. By using an artificial neural network, robot developers can train the robot through these scenarios by giving it a goal and punishing or rewarding it for things like repeatedly bumping into the same object — essentially giving it the equivalent of thousands of lifetimes of experience.
According to a former DeepMind analyst, the advent of robotics simulators has not only made everyday robots more of a reality, but has also provided huge cost and time savings: the transition to simulation saves millions of dollars per robot development, and can save over 25 years of experimentation time.
Robot Vacuum Cleaners Are The Tip of the Iceberg — the Real Opportunity is in Military and Defense
The Roomba from iRobot was largely funded by DARPA and other military defense government agencies for development of a robot to identify and dispose of bombs in Iraq and Afghanistan. The military has invested heavily in robots that can function on rough terrain and in dark and compact spaces. These robots need a high level of spatial awareness to be able to perform defense tasks in complex and highly cluttered environments.
The urgent government need for spatially skilled robots has led to massive advances from Boston Dynamics and iRobot, among other major players. According to the engineer from Boston Robotics, drones and grounded robots that combine navigation, mobility, manipulation, and artificial intelligence have saved hundreds of American lives in Iraq and Afghanistan alone.
“We have a phrase for consumer-facing products in engineering: keep it simple stupid (KISS),” explained the Boston Dynamics engineer. “But the military is a different ball game. There’s no barrier to adoption, they just want a robot that will save lives and that actually works.”
The military’s interest in grounded robots and drones will lead to technological advances that will eventually enter into consumer products. The combination of deep learning and simulations has already contributed to incredibly efficient paths to robot development. Similar to how the Roomba came from a government contract for bomb disabling robots overseas, we will see the development of other consumer facing robots increase exponentially as the investments in research and development from the government pay off.