Evolving Morphological Adaption Methods in Compliant Robots

  • Katt E Walker

Student thesis: Doctoral ThesisDoctor of Philosophy (PhD)


Despite the huge success of robotics in general, there are very few machines that are capable to stably locomote in rough, unknown terrain. Whilst a few robots are just becoming able to deal with the uncertainty, complexity and variety typical of natural environments, in the most cases even these cutting edge robots fail.
On the other hand, animals, including humans, have the ability to locomote in many environments un-accessible to robots. Additionally, animals are able to outperform robots in almost any category (with respect to locomotion), including in energy efficiency, stability, robustness, agility, and numerous others. There are many possible reasons as to why nature is able to outperform robotic designs. It is speculated that animals are able to successfully locomote without an exact environmental model by outsourcing some of this computation to their well designed morphology, achieved through evolution. Additionally, if adaption to new environments is required, animals not only adapt their behaviour but also in some cases their morphology. They appear to be able to learn from their interaction with the environment and use this information to adapt. Whilst many researchers have improved robot morphology through artificial evolution, recreating adaptive morphology is relatively unexplored. This leads to the question, can artificial evolution be used to find optimal methods of morphological adaption by exploiting feedback from the environment in order to successfully locomote in a wide range of environments?

In an attempt to partly answer this question, this thesis forms two parts. Firstly, the best methods to adapt a simulated Spring Loaded Inverted Pendulum Model (the SLIP model). The SLIP model has a unique property that if the combination of its morphological and control parameters are in a particular range it is self stabilising. The aim of the first part of the thesis is to find ways that the SLIP model can adapt its parameters accordingly to become stable, based on its interaction with the environment. Two approaches are explored; an offline approach, where adaption of the SLIP model occurs between episodes and the model is allowed to fail after each episode and an online approach where instead the adaption takes place between strides. Not only did both methods expand the range of parameters for which self-stability could occur (when compared with a basic SLIP model that has no capacity for learning) but in the case of the online learning the model was now able to withstand environmental changes, such as a decrease in ground level of up to 14 times the length of the spring.

In the second part of the thesis, the optimal morphological adaption of a soft robot is evolved. The evolved morphological adaption method use the distribution of kinetic energy throughout the entire robot to determine which parts to harden and which to soften. If part of the robot becomes too soft, it is removed. Thus an optimal morphology, adapted for the specific environment is sculpted. Here, the results show that as the kinetic energy of the different robot parts depends on the interaction the entire robot has with the each environment, many task-specific final morphologies can be sculpted but crucially they are all created using one single general method of adaption.

Overall, it is hoped that the research presented in this thesis showcases the potential that artificial evolution has for enabling robots to learn to adapt their morphology, based on interaction with the environment, in order to achieve robust locomotion in a wide range of environments.
Date of Award23 Mar 2021
Original languageEnglish
Awarding Institution
  • University of Bristol
SupervisorHelmut Hauser (Supervisor) & Sabine Hauert (Supervisor)

Cite this