BRISA Reading Group 05 - Embodied Intelligence via Learning and Evolution

In our very first session, Play2Perfect had a robot learn to play before it learned to assemble and the discussion turned to evolution and how it relates to reinforcement learning. This time, we follow that thread directly. Most robot learning treats the body as fixed and asks how to learn a controller for it. This paper makes the body part of the search. An outer evolutionary loop mutates simulated creatures by adding, deleting and reshaping their limbs and joints. An inner RL loop then trains a controller for each new body from scratch. How well that controller performs at the end of training decides whether the body gets passed on. Nothing learned is inherited. Every child starts with a randomly initialised controller, and nothing in the setup rewards learning quickly. Even so, within ten generations, the training time needed to reach a fixed level of performance drops by half. There's also a very practical side. The full run used 1,152 CPUs and finished in under 16 hours. One small design choice, removing the oldest agent instead of the weakest, lets the system survive failing machines, so it can run on cheap spare cloud capacity. And a broader one: should we keep designing robot bodies by hand and learn only the controller, or should the body be learned too? No preparation needed. Skim the abstract if you have ten minutes. Everyone is welcome - whatever you study, wherever you study and however much robotics you have done so far. Come with questions, disagreements, or your own take on how much of intelligence lives in the body. 🎟️ There are only a limited number of spots, so we can keep the session personal and interactive. Please register early and only attend if you've received a ticket. 📍 Hosted at the Merantix AI Campus on September 22nd. Doors open 18:30, we start at 18:45. 👋 New here? Join the community group — everything else gets announced there first.