Neuro-evolutionary robotics is considered to be an impressive method to achieve the collective behaviors of swarms of robots. Several studies have investigated this area and various methods and concepts have been suggested but there have been very few empirical evaluations and comparative analyses.
Image Credit: Shutterstock.com/ sdecoret
A new study by Mauro Birattari and his research group from the research center IRIDIA, ÉcolePolytechnique de Bruxelles, UniversitéLibre de Bruxelles, has compared some of the most familiar and latest neuro-evolutionary techniques for the offline design of robot swarms.
Concretely, these methods can enable the development of humanoid robot behavior, but to my knowledge, neuro-evolutionary robotics is not yet routinely adopted in real-world applications.
Mauro Birattari, École Polytechnique de Bruxelles, UniversitéLibre de Bruxelles
All such processes involve using evolutionary algorithms to produce a neural network capable of controlling the robots. In other words, a neural network executes this process by considering the sensor readings as input and delivers actuator commands.
Computer simulations are employed to produce a neural network that is relevant for the particular mission that must be fulfilled by the robots. As soon as the neural network is produced (in simulation), it is installed on the physical robots and subjected to tests.
The researchers compared several methods and noticed a kind of “overfitting.” That is, the design process turns highly specialized in the simulation setting and the generated neural network fails to “generalize” to the real world. This difference between reality and simulation utilized in the design process is known as the reality gap. Despite the fair accuracy delivered by the simulator, the differences are unavoidable.
For example, if robots need to move back and forth between two areas, one solution that the evolutionary process might find in simulation is to produce a neural network that makes the robot move along a circular path that touches both areas. This solution is very elegant and works very efficiently in simulation.
Mauro Birattari, École Polytechnique de Bruxelles, UniversitéLibre de Bruxelles
“When applied to robots, this solution would fail miserably: for example, if the real diameter of (one of) the robot’s wheels differs slightly from the nominal value, the radius of the trajectory will be different. The trajectory will no longer pass through the two given zones as desired and as predicted by the simulation,” added Birattari.
The Chocolate Solution?
Despite being counter-intuitive, the solution appears to be to decrease the “power” of the design method, that is, to adopt a technique capable of producing a limited range of behaviors. This implies that the researchers may get the worst results in simulation and be ready to accept it.
This method will not deliver as well in simulation as a “powerful” method due to its inability to exploit all the characteristics of the simulator. However, the results will be more general, less specialized to the simulator, and thus could be more likely to generalize to reality. It becomes better by being simpler.
The chocolate method seems to be a good demonstration of this concept. The chocolate process was suggested by researchers at IRIDIA Center a few years ago and does not belong to neuro-evolutionary robotics. However, it operates like the neuro-evolution, which automatically produces control software for robots, by operating under the same conditions.
The chocolate process operates on pre-existing software modules which exhibit low-level behavior (for example, I go in the direction of the light, I stop, I move away from perceived peers) and conditions for moving from one low-level behavior to another (for example, I am surrounded by peers, the color of the floor I am on is black.).
Chocolate prefers to play with predefined building blocks that are relatively more “coarse” rather than playing with a highly robust neural network that can produce an extended range of behaviors. The postulation for operation is that by behaving so, the risk of “over-fitting” will tend to decrease.
Journal Reference:
Hasselmann, K., et al. (2021) Empirical assessment and comparison of neuro-evolutionary methods for the automatic off-line design of robot swarms. Nature Communications. doi.org/10.1038/s41467-021-24642-3.