Imagine a scientist who wants to send a robot to explore in a faraway place, a place whose geography might be completely unknown and perhaps inhospitable. Now imagine that instead of first designing that robot and sending it off in the hope that it might be suitable, instead, she sends a robot-producing technology that figures out what kind of robot is needed once it arrives, builds it and then enables it to continue to evolve to adapt to its new surroundings.
It’s exactly what my collaborators and I are working on: a radical new technology which enables robots to be created, reproduce and evolve over long periods of time, a technology where robot design and fabrication becomes a task for machines rather than humans.
Robots are already all around us, in factories, in hospitals, in our home. But from an engineer's perspective, designing a shelf-stacking robot or a Roomba to clean our home is relatively straightforward. We know exactly what they need to do, and we can imagine the kind of situations they might find themselves in. So we design with this in mind. But what if we want to send that robot to operate in a place that we have little or even no knowledge about? For example, cleaning up legacy waste inside a nuclear reactor where it's unsafe to send humans, mining for minerals deep in a trench at the bottom of the ocean, or exploring a faraway asteroid. How frustrating would it be if the human-designed robot, that had taken years to get to the asteroid suddenly found it needed to drill a hole to collect a sample or clamber up a cliff but it didn't have the right tools or the right means of locomotion to do so? If instead we had a technology that enabled the robots to be designed and optimized in situ, in the environment in which they need to live and work, then we could potentially save years of wasted effort and produce robots that are uniquely adapted to the environments that they find themselves in.
So to realize this technology, we've been turning to nature for help. All around us, we see examples of biological species that have evolved smart adaptations that enable them to thrive in a given environment. For example, in the Cuban rainforest, we find vines that have evolved leaves that are shaped like human-designed satellite dishes. These leaves direct bats to their flowers by amplifying the signals that the bats send out, therefore, improving pollination. What if we could create an artificial version of evolution that would enable robots to evolve in a similar manner as biological organisms?
I'm not talking about biomimicry, a technology which simply copies what's observed in nature. What we're hoping to harness is the creativity of evolution, to discover designs that are not observed here on Earth, the human engineer might not have thought of or even be capable of conceiving. In theory, this evolutionary design technology could operate completely autonomously in a faraway place. But equally it could be guided by humans. Just as we breed plants for qualities such as drought resistance or taste, the human robot breeder could guide artificial evolution to producing robots with specific qualities. For example, the ability to squeeze through a narrow gap or perhaps operate at low energy.
This idea of artificial evolution imitating biological evolution using a computer program to breed better and better solutions to problems over time isn't actually new. In fact, artificial evolution, algorithms operating inside a computer, have been used to design everything from tables to turbine blades. Back in 2006, NASA even sent a satellite into space with a communication antenna that had been designed by artificial evolution.
But evolving robots is actually much harder than evolving passive objects such as tables, because robots need brains as well as bodies in order to make sense of the information in the world around them and translate that into appropriate behaviors. So how do we do it? Surprisingly, evolution only needs three ingredients: a population of individuals which exhibit some physical variations; a method of reproduction in which offspring inherit some traits from their parents and occasionally acquire new ones via mutation; and finally, a means of natural selection. So we can replicate these three ingredients to evolve robots using a mixture of hardware and software. The first task is to design a digital version of DNA. That is a digital blueprint that describes the robot's brain, its body, its sensory mechanisms and its means of locomotion. Using a randomly generated set of these blueprints, we can create an initial population of 10 or more robots to kick-start this evolutionary process. We've designed a technology that can take the digital blueprint and turn it into a physical robot without any need for human assistance. For example, it uses a 3D printer to print the skeleton of the robot and then an automated assembly arm like you might find in a factory to add any electronics and moving parts, including a small computer that acts as a brain. And to enable this brain to adapt to the new body of the robot, we send every robot produced to an equivalent of a kindergarten, a place where the newborn robot can refine its motor skills almost like a small child would. To mimic natural selection, we score these robots on the ability to conduct a task. And then we use these scores to selectively decide which robots get to reproduce. The reproduction mechanism mixes the digital DNA of the chosen parent robots to create a new blueprint for a child robot that inherits some of the characteristics from its parents but occasionally also exhibits some new ones. And by repeating the cycle of selection and reproduction over and over again, we hope that we can breed successive generations of robots where, just like is often observed in biological evolution, each generation gets better than the last, with the robots gradually optimizing their form and their behavior to the task and the environment that they find themselves in.
Now, although this can all take place in a time frame that's much faster than biological evolution, which sometimes takes thousands of years, it's still relatively slow in terms of the time frames we might expect in our modern world to design and produce an artifact. It's mainly due to the 3D printing process, which can take more than four hours per robot, depending on the complexity and the shape of the robot. But we can give our artificial evolutionary process a helping hand to reduce the number of physical robots that we actually need to make. We create a digital copy of every robot produced inside a simulation in a computer, and we allow this virtual population of robots to evolve. Now it's quite likely that the simulation isn't a very accurate representation of the real world. But it has an advantage that it enables models of robots to be created and tested in seconds rather than hours. So using the simulator technology, we can quickly explore the potential of a wide range of robot types of different shapes and sizes, of different sensory configurations, and quickly get a rough estimate of how useful each robot may be before we physically make it. And we predict that by allowing a novel form of breeding in which a physical robot can breed with one of its virtual cousins, then the useful traits that have been discovered in simulation will quickly spread into the physical robot population, where they can be further refined in situ.
It might sound like science fiction, but actually there's a serious point. While we expect the technology that I've just described to be useful in designing robots, for example, to work in situations where it's unsafe to send humans or to help us pursue our scientific quest for exoplanetary exploration, there are some more pragmatic reasons why we should consider artificial evolution. As climate change gathers pace, it is clear that we need a radical rethink to our approach to robotic design here on Earth in order to reduce that ecological footprint. For example, creating new designs of robot built from sustainable materials that operate at low energy, that are repairable and recyclable. It's quite likely that this new generation of robots won't look anything like the robots that we see around us today, but that's exactly why artificial evolution might help. Discovering novel designs by processes that are unfettered by the constraints that our own understanding of engineering science imposes on the design process.
Thank you.
(Applause)
