Continual learning — the ability of a model to accumulate past experience and transfer it to new tasks — has been studied extensively, but mostly outside embodied settings. A new research paper introduces EmbodiedCLBench, presented as the first benchmark in embodied simulation designed to evaluate the continual learning capability of self-evolving agents.
The benchmark is built around compositional generalization as its core evaluation principle. It tests whether agents can first learn from basic tasks and then recombine the skills they acquired to solve complex, previously unseen advanced tasks.
Using EmbodiedCLBench, the authors ran extensive experiments across multiple agent harnesses and learning methods. The results revealed consistent limitations as well as distinctive behaviors across the tested systems.
According to the findings, advanced tasks are hardly solvable in a zero-shot setting, but experience gained from basic tasks produces consistent improvement. This benefit remained stable even as task complexity increased.
However, the improvement did not scale with the number of learning samples, and current agents were unable to extract meaningful information from additional experience. The authors conclude that continual learning remains a fundamental yet underexplored capability, and position their benchmark as a resource for designing effective self-evolving agents in embodied environments.
The code and dataset are available at the project homepage.