Worms, boring animals, inhabited the primordial shallow seas. For billions of years, they hoped to bump into microscopic food as they floated helplessly along.
Just ten million years later, the seas teemed with attacking trilobites, battle-hardened mollusks and burrowing snakes. This ‘Cambrian Explosion‘ rocked the biosphere in an instant of mass evolutionary disruption.
Fast-forward about 500 million years. A DARPA scientist argues, in the prestigious Journal of Economic Perspectives, that robots are ready for their Cambrian moment: a like explosion of body types and strategies. The article is a d4e favorite for its big-picture, evolutionary perspective. It even mentions a virtuous circle — aka the positive feedback loop, the hallmark of emergent behavior. That said, it left something out, something ecological. Its absence is important enough to cast into doubt the articles specific thesis of robotic disruption.
Our intent here is not to rule out a Cambrian-like event. There are factors setting up to create one. The direction researchers like Jeff Clune are taking seems especially promising. However, like in any complex system, path dependence rules: the first steps define the success of the journey. It is these first steps that deserve scrutiny.
Gil Pratt, the DARPA scientist, offers up two key drivers of a rapid robot evolution. First, the cloud allows robots to collect and share new learning. Whatever one robot experiences, every other robot remembers. What wonderful mimicry of human culture! Pratt adds to this cloud-based collective memory the possibility that robots will some day store and share the products of simulations, what Pratt calls their ‘imagination.’
Second, computing power and neural nets allow robots to codify and search through millions of existing algorithms, each optimal for making a decision under a given situation: if this, then that. This is a decent analogy for the human subconscious mind, a vast network of perceptual associations that we use to instantly nurture a crying baby or evade a snake.
Combine ‘cloud robotics’ with ‘deep learning,’ Pratt argues, and stand back as robot speciation goes exponential.
So what is wrong with Pratt’s vision?
What is missing is a sense of how adaptation and selection occurred in the Cambrian. Scientists postulate many causes for that evolutionary riot. We focus on one: the development of vision. Think of the proto-eye — a photon sensor — as an ‘enabling technology’ similar to machine learning for robots. Worms could now sense light energy reflected off their food. Mutations that allowed movement in that direction were selected for, as were tools of predation such as jaws. This in turn spurred strategies of defense: external armor and camouflage. As these improved, so did vision and predation tools, and this spurred even more defensive innovation. The mutations themselves occurred randomly, but the self-reinforcing feedback loop of innovation was not random. It is a type of order orchestrated by networks.
The positive Cambrian feedback loop required two things of mutations: autonomy and diversity. Pratt’s robot ecosystem may have neither. Its enabler, the cloud, forces a robot to give up autonomy in order to access shared learning. This means that a robot’s successful ‘mutation’ gets uploaded and disseminated immediately. This is fine, just fine, for making better and better robots. Optimization, though, is not the same as the emergence of completely new strategies that result from a competitive ecological dynamic. In other words, if the robots aren’t trying to murder each other for food, it’s hard to see an explosive Cambrian-style innovation riot occurring.
Of course, different robot companies might fence off their collective robot cloud. This introduces the prospect of competition between their robots to fill niches like ‘housekeeping.’ Still, not as dynamic as predator-prey dynamics. (Imagine nano-robot ‘dust’ evading robot vacuums!) Even this level of competition is in question, however. Pratt mentions that household names such as Apple, Google and Amazon are investing in the robot space. These are not pre-Cambrian senseless worms. They could possibly dominate the robot ecology, and the resulting concentration would lead to even less competition and therefore less innovation.
To match the Cambrian dynamic would require a truly level playing field with low cost of entry (mutation available to all) and fierce, competitive selection pressure. This is where Pratt’s employer emerges as a factor, particularly with d4e principles in mind. DARPA designed the original internet, wittingly or not, for emergence. They built out infrastructure, set communications standards, and sat back as others competed to innovate on the platform. The analogous move today would be for the government, via DARPA, to make an effort to design the robot ecology for emergence.
What would this require? For one thing, preventing the dominance of a few firms. Perhaps by constructing an open-access robotic cloud — or even multiple ones in competition with each other and uncombinable. Further, DARPA could offer massive prizes for agent-based robot simulations and actual robot competitions. AI researchers with an evolutionary approach could compete in these to win grant funding for their labs. These would be robot-on-robot battles as well as robot task ‘races.’ All of these are present today, but their scale needs to be much larger to defeat the evolutionary advantage of existing behemoths like Google. Further, they could pit “cloud-connected” robots against autonomous enemies, thus increasing the chances of a truly novel adaptation. The result would be both autonomy and diversity. The way to measure the scale of an effort needed is not that it be ‘big,’ but that it be enough to tip the system into competition and away from a rich-get-richer dynamic.
Perhaps DARPA needn’t get involved. Silicon Valley, after all, generates plenty of innovation and disruption on its own. Perhaps, but recent evidence is actually discouraging. Network dynamics tend to lead to ‘unicorns’ that quickly dominate niches: Facebook, for example. How much more innovation would there be if Facebook had 40 competitors fighting tooth and nail for billions of users? All would have big scale and plenty of cash to invest in technology. Such an ecology would arguably be much better for the rest of us than Facebook’s dominance. Of course, someday, a firm will dethrone Facebook with a better mousetrap.
Unfortunately, ‘someday’ does not quite match the pace of the Cambrian Explosion.
