University of Oslo Teaches Robots to Share Brains and Evolve Faster
Researchers show that soft robots trained with knowledge from peers learn faster than isolated learners, and learning from multiple teachers yields more consistent gains.
PRO- Social learning beats individual learning for co-evolving soft robot bodies and brains under equal compute.
- Bayesian optimization trains each brain; peers donate initial controller samples as warm starts.
- Learning from multiple teachers produces more consistent gains than learning from one.
- Morphological similarity did not beat picking the best-performing teachers in these experiments.
- With unlimited re-learning, all approaches converge, so bodies are equally capable.
- Code and data are public; four tasks tested on Evolution Gym.
Evolving a robot's body and its controller at the same time is one of the messier problems in robotics. Shape determines what kind of movement is possible, but you can only tell if a morphology is good once you have trained a decent controller for it, and training a controller is expensive. A new paper from the University of Oslo, University of Toulouse Capitole, and the University of Trieste, accepted at GECCO 2026, asks a simple question: if one robot already figured out how to move a squishy body, can others skip ahead by borrowing what it learned?
Their answer is yes, and the details are more interesting than the headline suggests.
Why body-brain co-evolution is painful
In evolutionary robotics, the search space is tangled because body and brain depend on each other, and a promising morphology often gets discarded before its controller has a chance to catch up. The usual fix is a nested loop: an outer evolutionary process tweaks bodies, and an inner learning process trains a controller for each new body. That works, but every offspring pays the full cost of learning from scratch.
Lamarckian inheritance, where offspring receive optimized control parameters from their parents, has improved performance in prior work. The authors push further and ask why knowledge transfer should be restricted to parent-child pairs when any robot in the population might have something useful to contribute.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
