EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents
The benchmark tests whether embodied agents can turn basic learned skills into unseen, more complex tasks.
EmbodiedCLBench is presented as the first embodied-simulation benchmark for continual learning in self-evolving agents. The authors center the evaluation on compositional generalization: agents learn basic tasks, then try to recombine those skills on advanced tasks they have not seen. Their experiments find that basic-task experience helps compared with zero-shot performance, even as task complexity rises. But adding more learning samples does not produce matching gains, suggesting current agents are not extracting much useful signal from extra experience. Source: HF Daily Papers' note.
EmbodiedCLBench is presented as the first embodied-simulation benchmark for continual learning in self-evolving agents. The authors center the evaluation on compositional generalization: agents learn basic tasks, then try to recombine those skills on advanced tasks they have not seen. Their experiments find that basic-task experience helps compared with zero-shot performance, even as task complexity rises. But adding more learning samples does not produce matching gains, suggesting current agents are not extracting much useful signal from extra experience. Source: HF Daily Papers' note.
score 4