Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evolving generalist controllers to handle a wide range of morphological variations

Corinna Triebold, Anil Yaman

TL;DR

The paper addresses the challenge of robustness and generalization in evolved ANN controllers under morphological variation. It introduces an algorithm that evolves generalist controllers by exposing a range of morphologies during training, using evolutionary branching to partition the morphology space and training schedules to influence the learning trajectory. The study demonstrates that generalists achieve better performance across unseen morphologies at the cost of some specialization on specific morphologies, and that incremental morphology introduction often yields the strongest generalization, outperforming a baseline in key tasks. These findings advance the understanding of robustness and generalization in morphologically variable agents and offer a computationally efficient approach to controller ensembles for diverse morphologies in robotics.

Abstract

Neuro-evolutionary methods have proven effective in addressing a wide range of tasks. However, the study of the robustness and generalizability of evolved artificial neural networks (ANNs) has remained limited. This has immense implications in the fields like robotics where such controllers are used in control tasks. Unexpected morphological or environmental changes during operation can risk failure if the ANN controllers are unable to handle these changes. This paper proposes an algorithm that aims to enhance the robustness and generalizability of the controllers. This is achieved by introducing morphological variations during the evolutionary training process. As a results, it is possible to discover generalist controllers that can handle a wide range of morphological variations sufficiently without the need of the information regarding their morphologies or adaptation of their parameters. We perform an extensive experimental analysis on simulation that demonstrates the trade-off between specialist and generalist controllers. The results show that generalists are able to control a range of morphological variations with a cost of underperforming on a specific morphology relative to a specialist. This research contributes to the field by addressing the limited understanding of robustness and generalizability and proposes a method by which to improve these properties.

Evolving generalist controllers to handle a wide range of morphological variations

TL;DR

The paper addresses the challenge of robustness and generalization in evolved ANN controllers under morphological variation. It introduces an algorithm that evolves generalist controllers by exposing a range of morphologies during training, using evolutionary branching to partition the morphology space and training schedules to influence the learning trajectory. The study demonstrates that generalists achieve better performance across unseen morphologies at the cost of some specialization on specific morphologies, and that incremental morphology introduction often yields the strongest generalization, outperforming a baseline in key tasks. These findings advance the understanding of robustness and generalization in morphologically variable agents and offer a computationally efficient approach to controller ensembles for diverse morphologies in robotics.

Abstract

Neuro-evolutionary methods have proven effective in addressing a wide range of tasks. However, the study of the robustness and generalizability of evolved artificial neural networks (ANNs) has remained limited. This has immense implications in the fields like robotics where such controllers are used in control tasks. Unexpected morphological or environmental changes during operation can risk failure if the ANN controllers are unable to handle these changes. This paper proposes an algorithm that aims to enhance the robustness and generalizability of the controllers. This is achieved by introducing morphological variations during the evolutionary training process. As a results, it is possible to discover generalist controllers that can handle a wide range of morphological variations sufficiently without the need of the information regarding their morphologies or adaptation of their parameters. We perform an extensive experimental analysis on simulation that demonstrates the trade-off between specialist and generalist controllers. The results show that generalists are able to control a range of morphological variations with a cost of underperforming on a specific morphology relative to a specialist. This research contributes to the field by addressing the limited understanding of robustness and generalizability and proposes a method by which to improve these properties.
Paper Structure (23 sections, 1 equation, 7 figures, 1 table, 1 algorithm)

This paper contains 23 sections, 1 equation, 7 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. Neuro-evolution for continuous control problems
  4. Robustness and Generalizability
  5. Evolutionary ensembles
  6. Evolutionary branching
  7. Evolving generalist controllers
  8. Experimental setup
  9. Morphological learning sets
  10. Experimental parameters
  11. Training schedules
  12. Evaluation metrics
  13. Results
  14. Problem Outcomes
  15. CartPole
  16. ...and 8 more sections

Figures (7)

  • Figure 1: CartPole, specialists excel on the default morphology but perform poorly on local and global variations compared to generalists.
  • Figure 2: Bipedal Walker, specialists excel on the default morphology but perform poorly on local and global variations compared to generalists.
  • Figure 3: Ant, specialists excel on the default morphology but perform poorly on other variations.
  • Figure 4: Walker2D, specialists excel on the default morphology but perform poorly on other variations.
  • Figure 5: Illustration of fitness progression of five independent evolutionary processes (on morphology set size of 64).
  • ...and 2 more figures