This benchmark evaluates continual learning for self-evolving embodied agents.
A new embodied benchmark tests continual learning by requiring agents to recombine basic skills for unseen advanced tasks, revealing limited scaling.
Key Takeaways
- Embodied CL Bench is the first benchmark for evaluating continual learning in self-evolving agents within embodied simulation.
- It tests compositional generalization: agents must learn basic skills and recombine them to solve unseen, more complex advanced tasks.
- Prior experience from basic tasks consistently improves performance on advanced tasks, even as task complexity increases.
- Current agents fail to scale this benefit with more learning samples, suggesting the bottleneck is converting experience into reusable skills.
- The benchmark separates two key questions in agent development: whether earlier learning helps, and whether more learning helps more.
Meta description: Embodied CL Bench tests whether self-evolving agents can learn from basic tasks and recombine acquired skills to solve complex, unseen advanced tasks.
Tags: Embodied CL Bench, continual learning, self-evolving agents, embodied simulation, compositional generalization
Featured image: an abstract, friendly illustration of a learning agent connecting basic task skills into advanced task sequences, with no real people, no logos, and no text.
Embodied CL Bench is a new benchmark for continual learning in embodied simulation. The authors introduce it as the first benchmark in embodied simulation for evaluating the continual learning capability of self-evolving agents. Its results point to a practical problem: agents can benefit from prior experience, but they do not reliably benefit more as experience grows.
Why embodied settings need a new test
Continual learning requires models to accumulate past experience and transfer it to new tasks. Much of the research in this area has focused on non-embodied settings, where learning is measured through task sequences, predictions, or dataset-based evaluation. That leaves a gap: when an agent acts inside a simulated environment, it must connect earlier interactions with later goals, observations, and action requirements.
Embodied CL Bench is designed to address that gap. Instead of asking only whether an agent can complete a familiar task, the benchmark asks whether earlier learning can be reused in an embodied environment for later, more difficult tasks. This makes it relevant to self-evolving agents, which are expected to update their behavior over time rather than respond only to a fixed set of inputs.
What compositional generalization checks
The benchmark uses compositional generalization as its core evaluation principle. In simple terms, this means agents must learn skills from basic tasks and then recombine those skills to solve complex, unseen advanced tasks.
That is a stricter test than memorization. An agent that only reproduces a successful procedure may fail when a later task requires learned pieces to be arranged in a new way. Embodied CL Bench therefore evaluates a two-stage process: first, agents learn from basic tasks; second, they apply the acquired skills to advanced tasks that require new combinations.
Where current agents still fall short
Using Embodied CL Bench, the researchers run extensive experiments across multiple agent harnesses and learning methods. The findings show consistent limitations and distinctive behaviors.
Under a zero-shot setting, advanced tasks can be hardly solved. When the agent is given the advanced task directly, without useful experience from the basic tasks, it struggles. Prior experience improves performance: experience from basic tasks enables consistent improvement on later advanced tasks.
The benchmark also shows that this benefit remains stable even when task complexity increases. In other words, harder advanced tasks do not automatically remove the advantage gained from earlier learning.
The main failure appears in scaling. The improvement does not scale accordingly with the number of learning samples. Current agents cannot extract meaningful information from additional experience. This suggests that the bottleneck is not simply giving agents more examples; the harder problem is turning experience into reusable skill.
Why the benchmark matters for agent development
The results highlight continual learning as a fundamental yet underexplored capability for self-evolving agents. For developers, Embodied CL Bench offers a way to separate two questions that are often mixed together: whether earlier learning helps at all, and whether more learning helps more.
For readers following agent research, the benchmark suggests that embodied self-evolution needs methods that can reorganize experience into transferable skills. A test built around compositional generalization can help reveal whether new harnesses or learning methods improve that reorganization, rather than only improving performance on isolated tasks.
The researchers provide the benchmark’s code and dataset at https://pku-value-lab.github.io/Embodied-CL-Bench-homepage. This gives a shared starting point for evaluating continual learning inside embodied simulation, while keeping the test tied to one concrete challenge: can an agent learn from basic tasks and then recombine what it learned to solve unseen advanced tasks?

Comments (0)