EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

EmbodiedCLBench. Evaluating Continual Learning for Self-Evolving Embodied Agents. By Jie Huang and eight others, from Peking University, the State Key Laboratory of General Artificial Intelligence at the Beijing Institute for General Artificial Intelligence, and one other institution. Published on the ninth of October. Its code repository is worked out: about two hundred source files, with a readme.

Continual learning asks a model to accumulate what it has experienced and carry it over to new tasks. There is substantial research on it, but almost all of it sits in non-embodied settings. Agents that act inside an environment, with a body and a world around them, have largely been left out. The idea of continual learning has not been extended to embodied environments, and without a way to measure it there, nobody can say how well self-evolving embodied agents actually accumulate and transfer what they learn. That is the gap this work sets out to fill.

EmbodiedCLBench is presented as the first benchmark in embodied simulation for measuring the continual learning ability of self-evolving agents. Its central principle is compositional generalization. The question it puts to an agent is whether it can first learn from basic tasks, and then recombine the skills it acquired to solve advanced tasks that are complex and that it has never seen.

The design follows that principle in two tiers. Basic tasks are where experience is gathered. Advanced tasks are where that experience is supposed to pay off, through recombination rather than repetition. On top of this, the authors run extensive experiments across several agent harnesses and several learning methods, so that the limits they find are not tied to a single setup. Testing on unseen advanced tasks is what keeps the measurement honest: an agent cannot pass by memorizing what it practised, only by recombining it.

Several patterns hold consistently. Under a zero-shot setting, the advanced tasks can hardly be solved at all. Experience from the basic tasks changes that, producing consistent improvement. That benefit is also stable: it holds up even as task complexity increases.

The limitation is in scaling. The improvement does not grow in step with the number of learning samples. Giving agents more experience does not help them further, because current agents cannot extract meaningful information from the additional experience. Learning from a little works; learning from more, so far, does not add much.

The authors read these results as a sign that continual learning is a fundamental capability that remains underexplored for embodied agents. The benchmark, with its code and dataset released, is offered as a resource for building self-evolving agents that can keep learning inside embodied environments, and especially for working on the part that does not yet scale.