What is EvoSphere? EvoSphere is a concept that captures the world of robot evolution as developed by Gusz Eiben and his research group on the Vrije Universiteit Amsterdam. A visualisation by an artist based on soft bodied robots is presented here; the technological details are discussed in an article that you can read here. EvoSphere represents a tangible design template of a habitat for evolving robot populations. It identifies the principal components of such evolving systems, thus it provides a basis for a real-world implementation.
How does it work? The main components of the system are a (re)production facility dubbed the Birth Clinic, a training arena called the Nursery, and the Arena that represents the world where the robots operate. New robots are created in the Birth Clinic using an ‘organ bank’ of prefabricated components (e.g., CPUs, light sensors, servo motors), 3D-printers, and an assembly facility. Newborn robots start their ‘lives’ in the Nursery. Here they learn to control their own body (which may be different from the bodies of the parents) and to perform basic tasks offered in a systematic order and monitored by a camera + computer system and/or a human user. If a robot acquires the required set of skills it is declared an adult and enters the Arena where it must survive, reproduce, and perform user-defined tasks; otherwise it is removed and recycled. The Nursery increases the chances of success in the Arena and plays an important evolutionary role: it prevents reproduction of poorly performing robots. Lifetime learning continues in the Arena, but now without centralized supervision, autonomously. Reproduction is driven by a mate selection mechanism (innate to the robots or executed by the human breeder) to identify two or more robots for parenting a child. Judging a possible mating partner can be based on measurable task performance or robotic versions of `attractiveness’. If they evaluate each other favourably, the parents transmit their own genomes to the Birth Clinic where these undergo crossover and mutation and the resulting new genome is used to seed the construction of a child robot.
Does it really exist? Currently, no. There are computer simulations that have many of the main EvoSphere features, but no physical implementations. The Robot Baby Project is the first attempt to build a simplified implementation as a proof of concept.
What can you do with it? It depends on the type and size of the specific (future) implementation. A laboratory version can be used as a research tool for robotics, artificial intelligence, and (evolutionary) biology. An bigger outdoor version could be a robot breeding farm, where users select those robots to reproduce that perform the given tasks best. After evolving a good robot design the breeding process can be stopped and the `winning’ robot can be produced on a large scale for deployment in real task environments. These could be challenging environments, where it is hard to determine the optimal form of robots by theoretical models, for instance, forests. An EvoSphere could also be the core of entire robotic ecosystems that evolve and work for long periods in challenging environments without the need for direct human oversight. Such systems can do, for instance, ore extraction on the seafloor or exploration of other planets.
What about safety? The EvoSphere concept is designed with a centralized reproduction facility, the Birth Clinic, for reasons of safety. Evolutionary robots in the future could self-reproduce without humans-in-the-loop. Hence, distributed reproduction mechanisms (e.g., self-assembly or robotic equivalents of eggs or pregnancy) should be avoided and the system should have a central ‘kill switch’ to stop reproduction if necessary. The Birth Clinic is such a switch; shutting it down stops robot reproduction.
What? The Robot Baby Project (VU Amsterdam, 2015-2016) is a focused attempt to demonstrate that robots can have children. The scientific background is provided by a model of robotic reproduction and evolution, published the 2013 paper called: The Triangle of Life: Evolving Robots in Real-time and Real-space. The Triangle of Life framework describes the pivotal life cycle of self-reproducing robots. This life cycle does not run from birth to death, but from conception (being conceived) to conception (conceiving one or more children) and it is repeated over and over again, thus creating consecutive generations of robots. The result is an evolving population of robotic organisms, where the bodies as well as the brains can adapt to the given environment. The Triangle of Life is an abstract model, EvoSphere -discussed on the previous tab- is a tangible incarnation of it.
Why? The robot baby project is a proof of concept. Its main objective is to implement all three constituents of the Triangle of Life in a simplified form and to connect the dots, that is, complete one full life cycle. It is to prove the feasibility of robots that can reproduce in hardware, in the real world, rather than in software simulation. With this demonstration we hope to initiate a healthy scientific discussion and inspire further research.
How? The main premise is simplification. Each system component is made as simple as possible in order to be able to integrate them into a full cycle. In particular, we have:
- Specified a certain type of robots (bodies and brains) and the genetic language that can describe a robot through an artificial genotype (DNA). Our design is based on RoboGen.
- Set up a procedure that starts with a genotype (code for a certain robot) and ends with a phenotype, a physical robot designated by the given genotype. This amounts to the birth or morphogenesis process in the Triangle.
- Implemented learning method for infant robots to learn to control their own body. This belongs to the infancy stage.
- Established a reproduction mechanism that regulates mate selection and randomised recombination of the parental genomes. This implements adult life.
For the specific demonstration we have constructed two robots. We let them go through the infancy stage and become adults. The skill they had to learn was locomotion and navigation to a specific spot, the `mating corner’ of the habitat. Once they met in the mating corner, they mated (virtually) and sent their DNA to the Birth Clinic. This consisted of a 3D printer and collection of `body parts’, such as CPUs, light sensors, servo motors. The parental genotypes were randomly recombined into a new piece of DNA and a new robot was printed and assembled according to this specification. This delivered the first robot baby and concluded the first robotic life cycle. The project achieved its objective. We have gained much know-how and could identify important issues for further research and development.
Our robots are based on the RoboGen system, developed at the EPFL. We are indebted to Josh Auerbach for his collaboration in this project.
Simplification being our main premise, we limit the set of body parts to a minimum. We use three types of 3D-printable components: a head block, a body block, and a joint. To obtain a functional robot we combine these with not printable parts, such as servos to drive the joints, light sensors to be the ‘eyes’, a Raspberry Pi to be the ‘brain’ , and a rechargeable battery to provide energy. These components and their possible combinations determine all realizable robot bodies that can be created within our system.
We decided to make the first parent with the shape of a ‘spider’, be it with fewer legs. It has a central module with the CPU, the battery, and a light sensor in the middle and four limbs with two blocks and two servo driven joints. The colour of the 3D-printed blocks is blue.
The second parent has the shape of a ‘gecko’. Its head, the central module with the CPU, the battery, and a light sensor, is in front. Its body shows a left-right symmetry, the limbs are shorter than the spider’s. The colour of the 3D-printed blocks is green.
The child is a randomised recombination of the two parents. Its colours show the origin of certain body parts; the white colour of the head indicates a mutation.
Topics addressed here include: artificial intelligence, evolutionary computing, evolutionary robotics, evolution of intelligence, evolution of morphology, Evolution of Things, EvoSphere, robot baby project, the Triangle of Life.
This website has been set up by Artificial Intelligence researchers of the Vrije Universiteit Amsterdam. It is to provide information about ongoing work concerning the evolution of intelligence in robot populations.