Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

More complex environments may be required to discover benefits of lifetime learning in evolving robots

Ege de Bruin, Kyrre Glette, Kai Olav Ellefsen

TL;DR

The paper investigates whether intra-life learning enhances evolving robot locomotion by adding a Bayesian Optimization-based control-learning phase to morphology evolution and comparing performance in flat versus hills environments. Using a Revolve2 modular framework and MuJoCo simulation, it tests learning budgets of 1, 30, and 50 iterations to assess how control optimization interacts with environmental complexity. The key finding is that lifetime learning yields more pronounced gains in the hills environment, with significant performance differences observed after substantial function evaluations, underscoring the importance of environmental difficulty for observing lifelong learning benefits. The work suggests that environment design is crucial for evaluating learning-enabled evolution and points to future directions including no-learning baselines and Lamarckian extensions to control.

Abstract

It is well known that intra-life learning, defined as an additional controller optimization loop, is beneficial for evolving robot morphologies for locomotion. In this work, we investigate this further by comparing it in two different environments: an easy flat environment and a more challenging hills environment. We show that learning is significantly more beneficial in a hilly environment than in a flat environment and that it might be needed to evaluate robots in a more challenging environment to see the benefits of learning.

More complex environments may be required to discover benefits of lifetime learning in evolving robots

TL;DR

The paper investigates whether intra-life learning enhances evolving robot locomotion by adding a Bayesian Optimization-based control-learning phase to morphology evolution and comparing performance in flat versus hills environments. Using a Revolve2 modular framework and MuJoCo simulation, it tests learning budgets of 1, 30, and 50 iterations to assess how control optimization interacts with environmental complexity. The key finding is that lifetime learning yields more pronounced gains in the hills environment, with significant performance differences observed after substantial function evaluations, underscoring the importance of environmental difficulty for observing lifelong learning benefits. The work suggests that environment design is crucial for evaluating learning-enabled evolution and points to future directions including no-learning baselines and Lamarckian extensions to control.

Abstract

It is well known that intra-life learning, defined as an additional controller optimization loop, is beneficial for evolving robot morphologies for locomotion. In this work, we investigate this further by comparing it in two different environments: an easy flat environment and a more challenging hills environment. We show that learning is significantly more beneficial in a hilly environment than in a flat environment and that it might be needed to evaluate robots in a more challenging environment to see the benefits of learning.

Paper Structure

This paper contains 8 sections, 1 equation, 3 figures.

Table of Contents

  1. Introduction
  2. Methods
  3. Phenotype
  4. Genotype
  5. Evolution and Learning
  6. Simulation
  7. Results
  8. Conclusion

Figures (3)

  • Figure 1: An example robot in the two environments. Left is the flat environment and right is the hills environment.
  • Figure 2: Plots comparing the fitness of different learning budgets, left plots on the flat environment and right plots on the hills environment. The line shows the mean fitness of the population after the number of morphologies evaluated/function evaluations, combining all runs for that learning budget. Shaded areas are the standard deviation.
  • Figure 3: Box plots showing the mean fitness of the generation after 100.000 function evaluations.