Collections of hundreds of physical cubic modules communicate only with neighbors to infer shape class and localize damage. Trained NCAs transfer from simulation to real hardware with strong robustness to faults and out-of-distribution variations.

Combining evolutionary Quality-Diversity search with language models to generate constraint-compliant architectural designs. Shows that QD-synthesized training data is crucial for text-guided generative design.

MarioGPT uses large language models to generate Super Mario Bros levels from text prompts, enabling controllable procedural content generation and open-ended discovery of diverse gameplay dynamics.

Skill Decision Transformer extends language model-based offline reinforcement learning by discovering diverse primitive behaviors from trajectory data. Using hindsight relabeling and skill discovery, it performs state-marginal matching and remains competitive with supervised offline RL approaches on D4RL through reward-free optimization.

3D Neural Cellular Automata that grow complex Minecraft structures from single cells and regenerate functional machines, demonstrating morphogenesis in volumetric environments.