IAAC – Master in Robotics and Advanced Construction
Workshop 3.1
Faculty: Raimund Krenmueller
Faculty Assistant: Soroush Garivani


Photo credits: Raimund Krenmüller


Thanks to increasing computational power and miniaturisation, together with the development of better and cheaper sensors, incorporating real-time data in the control of digital production processes is becoming commonplace, driving the transition from machines that execute rigid production sequences towards flexible behaviours, capable of operating in dynamic environments and handling material indeterminacies. 

Compared to biological examples shaped by evolution, those man-made systems and the exhibited complexity of their behaviours still falls short significantly. However, as the fidelity and speed of computer simulations are increasing, the power of simulated evolution is likely to increase as well, with the potential to yield behaviours of artificial agents (encoded as neural controllers) that are superior to what could be engineered in the classical sense. This is especially promising where artificial evolution could resolve the apparent paradox of expecting top-down objectives to emerge from individually controlled agents’ behaviours in a collective (think social insects)

In this workshop, neuro-evolution, the use of evolutionary algorithms to determine the defining parameters of artificial neural networks, is explored as a strategy to develop autonomous behaviours of machines (i.e. robots). A method and the associated software framework for artificial evolution of neural robot controllers using the NEAT family of algorithms is applied in a range of experimental setups with increasing complexity. The critical link between the digital and physical domain is established by transferring a neural network that was evolved in simulation to the control system of physical robots.

Alongside the technical implementation, a theoretical framework is necessary that addresses how, and in what form human design intent can be integrated in a construction system based on evolved autonomous behaviours. Where design is neither targeting a specific outcome, nor a specific process, but flexible and adaptive behaviours of autonomous production machines, implications for the role of the designer and questions of authorship need to be addressed.

Learning Objectives

– Understanding the core concepts of neuroevolution for robotics
– Develop the know-how and techniques to set up and run a neuroevolution process
– Develop the ability to use advanced robotic simulation tools and integrating them in a neuroevolution process
– Applying the acquired know-how in a scenario that is relevant for autonomous construction
– Deepening and applying knowledge about Neural Networks and Evolutionary Computation
– Deepening and applying knowledge about ROS for sensing and processing of real-time data
– Transferring neural controllers from simulation to reality, using ROS and the robots at IAAC (in the case of inaccessibility of the IAAC – due to covid 19, the workshop will focus on the development of digital simulation, keeping the physical testing with robots for a later stage).