Can both body and controller evolve?

Yes, if the representation and evaluation support both.

Must every candidate be built physically?

No. Many searches use simulation, but transfer to physical operation needs validation.

Does mutation always improve a robot?

No. Selection and evaluation determine whether a variation is useful under the chosen criteria.

AI Glossary / Reinforcement learning and robotics / Intermediate

Evolutionary robotics

Definition

Evolutionary robotics applies evolutionary search methods to robot controllers, physical designs, or both. Candidate solutions are evaluated, selected, and varied over successive generations to discover useful behavior or morphology. Evaluation may occur in simulation, on physical robots, or through a combination. The approach automates parts of the design search, but people still specify representations, evaluation conditions, constraints, and the criteria used to judge outcomes.

What it means

A designer can describe what counts as successful performance without specifying every detail of the controller that achieves it. An evolutionary algorithm then explores alternatives, preserving or favoring candidates according to its selection rules and generating new variations.
The word evolution refers to a computational process inspired by biological evolution. It does not imply that robots reproduce physically, become conscious, or improve without limits. Search results reflect the environment and criteria used during evaluation. A solution that scores highly under those conditions may perform poorly elsewhere.

How it works

Define what can vary. A candidate might encode neural-network weights, control parameters, sensor placement, body dimensions, or a combined body-and-controller design. The encoding determines which alternatives the search can reach.
Generate an initial population and evaluate candidates in specified tasks. A fitness function can combine measures such as successful completion, time, and energy, while constraints exclude unacceptable solutions. Other evolutionary methods can emphasize diversity or multiple objectives rather than a single scalar score.
Select candidates and create variations through operators such as mutation and, where appropriate, recombination. Repeat evaluation and selection according to a defined computational budget or stopping rule. Preserve information about random seeds, environments, and evaluation protocols so results can be checked.
Simulation can make many trials practical, but an optimizer may exploit details that do not match the physical world. This simulation-to-reality gap requires investigation through realistic variation, appropriate modeling, and physical validation. Evolutionary improvement also needs independent evaluation: repeatedly selecting on the same scenarios can favor narrow solutions that do not generalize.

Three practical examples

A robotics laboratory evolves controller parameters for locomotion and compares the resulting gait across different surfaces.
An engineering team explores sensor locations and control settings together to improve a defined inspection task.
A researcher uses evolved robot behavior to investigate hypotheses about adaptation, while distinguishing the artificial setup from biological evidence.

Think of it this way

Think of testing generations of prototype designs. Promising prototypes contribute features to the next round, and variations create new possibilities. The process can discover unexpected solutions, but it remains shaped by the workshop's materials, test course, and scoring rules. Winning on one test course is not the same as being best everywhere.

Worked example

Consider three invented controller candidates evaluated in simulation. Define fitness as successful deliveries minus twice the number of collisions, so F = D − 2C. Candidate A completes 18 deliveries with two collisions, giving 18 − 4 = 14. Candidate B completes 15 deliveries with no collisions, giving 15. Candidate C completes 20 deliveries with four collisions, giving 20 − 8 = 12.
Under this particular score, B ranks first, then A, then C. Suppose a mutation of B produces candidate D, with 17 deliveries and one collision. Its score is 17 − 2 = 15, tying B despite a different outcome profile.
Now change the collision penalty to five points. The scores become A = 8, B = 15, C = 0, and D = 12. B is now uniquely best among these candidates. This demonstrates that the evaluation rule can change selection even when the observed behavior is unchanged.
In a real application, a collision may need to be prohibited by a constraint rather than traded against productivity through a score. These invented simulation results illustrate arithmetic and objective design, not an acceptable collision policy. A mutation also has no guarantee of improvement; the new candidate must be evaluated.

Illustrative case study

In this fictional case, Alder Robotics searches for a controller for a small inspection platform. Early simulations favor a fast gait, but a physical trial reveals that the controller relies on traction assumptions that the real floor does not satisfy.
The team revises the evaluation environments, includes variation in relevant conditions, and reserves separate scenarios for checking generalization. It compares evolved candidates with a simple existing controller rather than assuming that the search method must produce a better solution.
The project report includes failed trials and computational effort alongside successful behavior. The final candidate remains subject to the same operating checks required of a manually designed controller.

Common misconceptions and mistakes

A frequent misconception is that evolution automatically finds a globally optimal design. Search is limited by representation, variation operators, evaluation budget, and the landscape created by the objective.
Another mistake is treating fitness as an objective measure of overall usefulness. It measures what the evaluation encodes and can reward unintended shortcuts. A high simulated score is not proof of physical reliability.
Evolutionary robotics also differs from a robot learning during deployment. Evolution can occur entirely before deployment, although evolutionary methods and within-lifetime learning can be combined. Physical self-reproduction is not required.

Business use

Businesses may use evolutionary robotics to explore complicated design spaces or generate alternatives that are difficult to specify directly. The practical case depends on evaluation cost, manufacturing constraints, reproducibility, and the ability to validate outcomes.
Compare the method with simpler tuning or engineering approaches under a fair budget. Report performance across representative conditions and separate the cost of discovering a candidate from the cost of running it after deployment.

Check your understanding

Which candidate wins with the original score?

Answer: B, among the initial three; D later ties it.

Why does changing the penalty alter the comparison?

Answer: Fitness combines outcomes according to selected priorities.

Why retain independent test scenarios?

Answer: They help reveal over-specialization to the conditions used during search.

Related terms

Evolutionary algorithms
Fitness function
Mutation
Morphology
Multi-objective optimization
evolutionary algorithm
neuroevolution
simulation-to-reality gap
embodied intelligence

Learn next

Learn next how candidate encodings and selection methods affect evolutionary search. Study simulation-to-reality transfer and robust evaluation before interpreting an evolved controller as suitable for physical deployment.

Sources

Nolfi and Floreano, Synthesis of Autonomous Robots Through Evolution, original review abstract: https://pubmed.ncbi.nlm.nih.gov/11849613/ Floreano and Urzelai, Evolutionary Robots with On-line Self-organization and Behavioral Fitness: https://www.sciencedirect.com/science/article/pii/S0893608000000320 Nygaard and colleagues, Real-World Evolution Adapts Robot Morphology and Control to Hardware Limitations: https://arxiv.org/abs/1805.03388

Previous: Evolutionary multimodal optimization Next: Evolutionary strategies