Editorial: Evolving Robotic Morphologies

David Howard, Kyrre Glette, Nick CheneyFrontiers Robotics AI 2022 · OpenReview

TL;DR

This editorial summarizes a special topic on evolving robotic morphologies, highlighting the growing importance of co-evolving a robot's physical form and its control system for enhanced adaptability. It showcases recent advancements in this field, emphasizing the potential for creating more robust and versatile robots capable of tackling complex real-world challenges.

What Problem Does This Paper Address?

The paper addresses the challenge of creating truly adaptable and resilient robots. Traditional robotics often focuses on optimizing a robot's control system for a fixed physical form. However, this editorial argues that considering and evolving a robot's morphology alongside its controller offers a richer design space, leading to systems that can better adapt to diverse and unforeseen environmental conditions. This approach is rooted in the concept of "embodied cognition," where the physical body and its interaction with the environment are integral to intelligent behavior. The authors note that a significant hurdle has been the difficulty in developing effective encodings for this coupled search space and the practical challenges of fabricating and testing morphologically adaptable hardware.

Methods

This editorial doesn't present a single method but rather summarizes the diverse approaches explored in the research articles within the special topic. These include:

  • Quality-Diversity (QD) Algorithms: Specifically, the paper mentions MAP-Elites as a QD algorithm that excels at finding diverse and high-performing solutions for morphology-controller pairings, outperforming traditional objective-based search algorithms.
  • Task-Agnostic Design: One approach focuses on automatically designing adaptable soft voxel-based morphologies in simulation, using the concept of "criticality" (a state near a phase transition) as an indicator of adaptability.
  • Epigenetic Error and Developmental Processes: The research explores how mechanisms like "epigenetic error" (mimicking transcription errors in gene expression) can introduce genetic variation without disrupting adapted genes, leading to increased adaptability.
  • Multi-Level Evolution (MLE): This paradigm decomposes the problem of modular robot evolution into different levels (material, component, robot), allowing for the discovery of reusable components and the generation of complex, high-performing robots.
  • Full-Cycle Reconfigurable Hardware: A platform called EMERGE is presented for evolving morphologies in full-cycle reconfigurable hardware, enabling the assembly, testing, and disassembly of modules in both simulation and real-world scenarios.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Embodied Cognition is Key: Evolving morphology and control together unlocks surprising and rich behaviors that emerge from the interplay between the robot's body, its controller, and its environment.
  • Morphological Diversity Matters: Quality-Diversity algorithms, like MAP-Elites, are crucial for generating a wide range of morphologies, which is beneficial for transferring robots to new and more challenging environments.
  • Encoding Biases are Critical: The way a robot's morphology and behavior are encoded significantly influences the attainable design space, highlighting the importance of understanding these biases.
  • Developmental Processes Enhance Evolution: Mechanisms mimicking biological development, such as epigenetic error, can provide valuable genetic variation for evolutionary processes.
  • Modularity and Reusability: Multi-level evolution approaches can lead to the creation of reusable robot components, streamlining the design of complex systems.
  • Real-World Embodiment is the Goal: While simulation is a powerful tool, the ultimate aim is to translate evolved morphologies and controllers to physical robots, with advancements in 3D printing and modular systems facilitating this.

Key Results

The editorial highlights the success of the research topic in bringing together 7 original research articles and one mini-review, involving seventeen authors. Specific results mentioned include:

  • MAP-Elites: Demonstrated to find higher-performing solutions and generate greater morphological diversity compared to objective-based search algorithms for morphology-controller pairings.
  • Hawthorne-Madell et al.: Showed that random developmental processes can increase genetic variation and adaptability in embodied computational evolution.
  • Chand and Howard: Presented a proof-of-concept for the Multi-Level Evolution (MLE) paradigm, demonstrating its ability to discover reusable robot components and generate designs difficult to achieve with conventional methods.
  • Moreno and Faiña: Successfully evolved 30 different morphologies in simulation and transferred them to real EMERGE modules, demonstrating the feasibility of real-world embodiment.

Limitations and Future Work

  • Challenges in Coupled Search Space: The paper acknowledges the inherent difficulty in navigating the expanded, coupled search space of morphology-plus-controller.
  • Fabrication and Assessment: Practical challenges remain in fabricating and assessing morphologically adaptable robotic hardware.
  • Sample Efficiency: The editorial emphasizes that sample-efficient evolution is a key prerequisite for real-world instantiation of physical evolving robotic systems.
  • Understanding Evolved Systems: Beyond building these systems, there's a need to develop the scientific understanding behind how they work.
  • Future Directions: The authors express hope that the special topic will inspire future research, with commonly noted features like soft robotics, quality-diversity, real-world evolution, and bioinspiration being promising avenues.

Summary updated

Abstract

Morphological adaptation is a key challenge for Evolutionary Robotics \citep{bongard2013evolutionary,doncieux2015evolutionary}, and robotics in general. Consideration of a robot’s physical form in addition to its control system holds particular promise for imbuing robotic systems with task and environmental adaptability \citep{sims1994evolving,lipson2000automatic}. The additional degrees of behavioural freedom that can be unlocked when morphology is considered in the evolutionary process allows access to a wide design space, providing a pathway towards resilient, rugged, and adaptable robots that can cope with unforeseen circumstances and be realistically deployed in real-world scenarios \citep{nygaard2021real}, with corresponding real-world impacts. Morphology evolution allows us to create and study systems based on the principles of embodied cognition, whereby coupling (evolved) controllers to (evolved) physical substrates in a meaningful environment \citep{miras2019effects} allows for surprising, rich behaviours that emerge from interactions between the three \citep{lehman2020surprising,lehman2011evolving}. Embodied cognition provides a route towards improving the complexity and adaptivity of evolved behaviours, leading to a corresponding increase in the applicability of evolutionary robotics systems to challenging scenarios.Systems that can fully exploit the principles of embodied cognition and automatically evolve environment-adapted and task-specific robots is an amb...
BibTeX
@inproceedings{howard2022editorial,
  author = {David Howard and Kyrre Glette and Nick Cheney},
  title = {{Editorial: Evolving Robotic Morphologies}},
  booktitle = {Frontiers Robotics AI},
  year = {2022},
  url = {https://openreview.net/forum?id=NpBgVwS-CiV},
}
Figure