The Overcooked Generalization Challenge (OGC) is a novel benchmark for reinforcement learning (RL) agents that evaluates their ability to collaborate with unknown partners in unfamiliar environments. Previous research has primarily evaluated collaborative RL with training environments or training partners, severely limiting our understanding of agents' generalization capabilities—a critical requirement for future collaboration with humans. OGC extends Overcooked-AI to support Dual Curriculum Design (DCD). It is fully GPU-accelerated, open-source, and integrated into the minimax DCD benchmark suite. Compared to previous DCD benchmarks, where designers manipulate only minimal elements of the environment, OGC provides a much richer design space, such as a full kitchen layout with multiple objects that requires consideration of the interaction dynamics between agents. OGC evaluates state-of-the-art DCD algorithms and scalable neural architectures, demonstrating that current methods fail to generate agents that generalize effectively to new layouts and unfamiliar partners. The results indicate that both agents and curriculum designers struggle with the joint challenge of partner and environment generalization. These results establish OGC as a challenging testbed for collaborative generalization and suggest key directions for future research. The code is open source.
