Source-linked AI summary
Evolving embodied intelligence from materials to machines
David Howard, Agoston E. Eiben, Danielle Frances Kennedy, Jean-Baptiste Mouret, Philip Valencia, Dave Winkler
TL;DR
Robots remain challenged by unstructured environments, where effective behaviour requires suitable combinations of body and brain and current body-design processes are constrained by manufacturing complexity. The paper proposes Multi-Level Evolution, which jointly evolves materials, components, and robot body plans for task-environment niches. It presents MLE as a route toward specialist robots that exploits advances in materials, manufacturing, and evolutionary robotics.
Problem
Robots struggle in complex, unpredictable environments, while robot body design has lagged because of manufacturing complexities despite the importance of body–brain combinations.
Method
Multi-Level Evolution searches materials, component geometries, and integrated robot body plans in a bottom-up evolutionary design process.
Results
MLE provides a proposed framework for generating task- and environment-specific robot specialists by harnessing diversity across materials, components, and body plans.
Takeaways & Limitations
MLE offers a research agenda for extending evolutionary robot design from conventional bodies and controllers to materials, components, and machines.
Abstract
from arXiv · showhide
Natural lifeforms specialise to their environmental niches across many levels; from low-level features such as DNA and proteins, through to higher-level artefacts including eyes, limbs, and overarching body plans. We propose Multi-Level Evolution (MLE), a bottom-up automatic process that designs robots across multiple levels and niches them to tasks and environmental conditions. MLE concurrently explores constituent molecular and material 'building blocks', as well as their possible assemblies into specialised morphological and sensorimotor configurations. MLE provides a route to fully harness a recent explosion in available candidate materials and ongoing advances in rapid manufacturing processes. We outline a feasible MLE architecture that realises this vision, highlight the main roadblocks and how they may be overcome, and show robotic applications to which MLE is particularly suited. By forming a research agenda to stimulate discussion between researchers in related fields, we hope to inspire the pursuit of multi-level robotic design all the way from material to machine.
The challenge of embodied intelligence
Robots struggle in unstructured environments because capable behaviour depends on tight body–brain–environment coupling, while robot body design has lagged behind manufacturing complexity. Multi-Level Evolution (MLE) addresses this gap by jointly searching materials, components, and robot body plans for task- and environment-specific specialists.
- The challenge of embodied intelligence: Embodied cognition links intelligent behaviour to the tight coupling of an agent’s body, brain, and environment rather than to the brain alone.The passage connects physical form and function to learning, development, and suitable in-environment behaviour.
- The challenge of embodied intelligence: Natural evolution varies across levels, from DNA and proteins to eyes, limbs, and body plans, motivating similarly free-form robot design.The paper presents multi-level variation as a route toward capable embodiments for challenging unstructured environments.
- The challenge of embodied intelligence: MLE searches from materials through components to integrated robot body plans, evaluating designs on tasks in environments while continually expanding candidate materials and components.The proposed three-level architecture discovers materials, assembles them into geometries, and integrates components into template body plans.
- The challenge of embodied intelligence: MLE contrasts with conventional engineering by generating specialists for task-environment niches through diversity across materials, components, and robot designs.The paper frames MLE as a universal designer spanning a wider design space than generalist-oriented engineering approaches.
- The challenge of embodied intelligence: The framework is intended to exploit abundant material possibilities and manufacturing advances while addressing the manufacturing complexity that has constrained robot body design.The paper identifies materials and manufacturing techniques as central to realizing more embodied robots.
Enabling technologies: advanced manufacturing
Advanced manufacturing can support MLE by producing free-form, multi-material structures and increasingly integrated robotic functions with less human intervention.
- Enabling technologies: advanced manufacturing: Multi-material additive and subtractive manufacturing can print intricate free-form components with complex geometries and in-situ material combinations.Functional gradation and voxel blending enable fine-resolution variation of properties such as stiffness and elasticity.
- Enabling technologies: advanced manufacturing: Printable sensors, actuators, and power systems are increasingly integrated into multifunctional materials, while semi-autonomous construction supports iterative generate-and-test processes.The paper describes a trajectory toward constructing whole robots without human intervention.
Inspiration and characteristics of Multi-Level architectures
MLE extends evolutionary robotics by searching materials, components, and robots as linked levels, producing diverse designs specialised to task–environment niches. Its architecture combines independent search processes, quality-diversity libraries, hierarchical pointers, and learned control.
- Inspiration and characteristics of Multi-Level architectures: MLE extends evolutionary robotics by adding materials and components to the search for task- and environment-specific robots.Classic evolutionary robotics typically searches controllers and body plans, while MLE incorporates real and newly discovered materials into a holistic design process.
- Inspiration and characteristics of Multi-Level architectures: The architecture uses vertically stacked robot, component, and material levels, with separate search processes that can target sensors, actuators, and structural elements.Components combine geometries with one or more materials, and each level can use suitable representations or algorithms.
- Inspiration and characteristics of Multi-Level architectures: Quality-diversity illumination fills multidimensional libraries with diverse, high-performing materials, components, and robots rather than seeking one optimum.Solutions are assigned to bins by physical properties, while fitness determines the best solution retained in each bin.
- Inspiration and characteristics of Multi-Level architectures: Lower levels bootstrap materials and assemble geometries into components, while the robot level combines components into body plans and evaluates mission performance.CPPNs define component geometries and body-plan layouts; pointers select specific entries from lower-level libraries.
- Inspiration and characteristics of Multi-Level architectures: MLE’s hierarchical search can enable emergent behaviours by exposing upper levels to diverse material–component combinations without prescribing how useful behaviour must arise.Robot fitness is evaluated in the environment, while component and material fitness can reflect cost or task-relevant properties.
- Inspiration and characteristics of Multi-Level architectures: For complex tasks, controllers such as neural networks, central pattern generators, behaviour trees, or modular architectures can direct body–environment interactions.Controllers may be optimised through reinforcement learning, evolutionary algorithms, imitation, or postdeployment online learning.
Physical and virtual testing provides the best of both worlds
MLE combines simulation, predictive modelling, and physical experimentation through a shared representation for real and virtual artefacts. This hybrid workflow can accelerate search while addressing the reality gap that limits simulator-designed robots.
- Physical and virtual testing provides the best of both worlds: Simulation and modelling are essential because manufacturing and testing every new material, component, and robot would be prohibitively costly and time-consuming.MLE shifts much of the search effort toward cheaper, parallelisable simulations and models.
- Physical and virtual testing provides the best of both worlds: Hybridisation gives real and virtual materials, components, and robots identical representations, allowing them to crossbreed and contribute candidates to either domain.Simulation can accelerate physical evolution, while favourable physical results can improve simulated evolution.
- Physical and virtual testing provides the best of both worlds: Physical experiments provide ground-truth data, but real-world robot evaluation remains difficult because of repeatability and physical-damage problems.Custom test arenas and generate-and-test facilities make physical evaluation increasingly feasible.
- Physical and virtual testing provides the best of both worlds: The reality gap can degrade performance when simulator-designed artefacts transfer to reality, so MLE uses real-design data, surrogate models, simulator calibration, and transferability functions.These techniques aim to improve simulator accuracy and speed across material, design, and controller levels.
- Physical and virtual testing provides the best of both worlds: MLE’s broader benefits include scalability through parallel level-specific searches, self-optimisation as models and libraries improve, and reuse of materials and components across architectures.These benefits are presented alongside collaboration across institutions with different hardware and specialist capabilities.
Opportunities for MLE architectures
MLE is especially suited to soft robotics and environmental niches where specialized combinations of materials, morphology, behavior, and degradation are needed. It offers an alternative to preconceived designs and costly manual specialization by automatically exploring niche-specific embodiments.
- Soft robotics: Soft robotics is a strong application because deformable robots integrate sensing, actuation, and structure, while MLE can search these functions jointly.A shared library of multifunctional components supports this integration more directly than separate sensing and actuation libraries.
- Soft robotics: MLE addresses soft robotics’ lack of a codified design methodology by exploring polymers, composites, and emergent body designs instead of copying preconceived biological forms.The approach asks what embodiments evolution might devise from artificial building blocks rather than designing an octopus- or jellyfish-like robot.
- Soft robotics: By discovering specialized materials and using them in emergent components, bodies, and controllers, MLE provides a pathway toward strongly embodied soft robots.The intended process links material discovery to increasingly capable componentry and control.
- Environmental monitoring: MLE can automatically design specialized robots for distinct environmental niches by combining materials, morphology, behavior, and degradation properties.Examples include wind-powered, sliding Antarctic robots; humidity-degrading Amazon crawlers; and solar-powered, heat-resistant Sahara robots.
Towards a new era of Embodied Intelligence
MLE is presented as a new robotic design paradigm that integrates materials, manufacturing, and multilevel evolutionary search to produce specialized embodiments. Its realization remains ambitious, requiring new representations, scalable architectures, staged development, and solutions to material, efficiency, and resource-allocation constraints.
- New design paradigm: MLE integrates different technologies and design levels under an evolutionary framework, potentially enabling radically new ways to produce robots.The authors frame this as a robotic design technique inspired by evolution’s ability to fill environmental niches with adapted lifeforms.
- Architectural development: A future MLE architecture could collapse sensing, actuation, and structural libraries into heterogeneous solution libraries, increasing emergence and integration but also computational and combinatorial costs.This alternate architecture is described as a further-future possibility rather than a current implementation.
- Challenges: Initial MLE designs will be limited to materials that are easy to create, characterize, and model.The authors expect high-throughput materials search and improved materials modelling to reduce this constraint over time.
- Representation: Hierarchical, level-spanning representations are identified as an opportunity for describing complex robots assembled from interacting artefacts across design levels.The paper contrasts this opportunity with the limited prior consideration given to hierarchical representation.
- Development path: The proposed development path progresses from laboratory prototypes with heavy human intervention, to constrained real-world missions, and eventually real-world deployments.The stages are projected over approximately five years, a decade, and twenty years, respectively.