Automating Design and Discovery: from bio-inspired robotics to biomimetic discovery

The INCOSE Nature Systems Working Group hosted a webinar on September 26/2014 by Hod Lipson on adaptive robotic systems that can handle unstructured or unforeseen circumstances without the control functions being 'designed in'.  Rather than emulating specific outcomes of evolution, Lipson's team is emulating evolution itself, starting with a collection of components, allowing connections to form and then combining/mutating/selecting the ones that perform the best.  The first attempts involved computer modeling and simulation, including various forms of 'soft robots' that evolved the ability to move.  Other projects involved physical robots where the controllers were able to evolve with the goal of robots that could jump, run and gallop. 

Lipson described three different approaches:

  • simulate robots and controllers in software (limited to simple robots)
  • evolution of physical robots (limited by the number of physical trials)
  • hybrid in which physical robots and the simulators co-evolved, significantly improving the speed of evolution (similar to co-evolution in nature)

Lipson described research on a self-modeling robot with four legs, eight motors (hip and knee joints) and two tilt sensors.  The robots initially make random motions.  By comparing self-directed actions with the outputs of the tilt sensors, the robots were able to develop various 'stick figure' models of itself.  The robot used these models to make predictions, test/evaluate the most divergent predictions and select/evolve the models.  The robot then looked for patterns of actuation that could move itself forward.  If a leg is damaged, the robot was able to modify its internal model because the model was no longer a good predictor of how the robot's actions interacted with the environment.  A movie showing both the learning and re-learning phases is available at  This adaptive modeling approach is not limited to robots but can be applied to modeling a black-box system through actively controlling inputs, observing the results and developing an internal model.  An example was a Cornell suspension bridge outfitted with vibrators and sensors which resulted in models that identified points of weakness faster and more accurately than traditional engineering methods.

Lipson also described an approach to evolving mathematical equations by comparing predicted to actual results.  The method involved co-evolution in that the 'evolution engine' repeatedly selected a subset of data points that resulted in the most disparate results from the various mathematical models.  The process has been successfully tested against known models and has been used to develop analytical models of biological systems (required that 'time delays' be added to the mathematical building blocks).  It can also be used to identify invariants, such as the total energy (kinetic plus potential) of simple and complex (double) pendulums.  The software is commercially available as Eureqa with 30-40K users (see a description at

Additional Reading:


Your rating: None