EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents
EmbodiedCLBench measures whether embodied agents can recombine basic skills for harder unseen tasks.
EmbodiedCLBench is a new embodied-simulation benchmark for continual learning in self-evolving agents, focused on compositional generalization. Agents learn basic tasks and must recombine those skills for complex unseen tasks. Across multiple harnesses, advanced tasks are rarely solved zero-shot; basic-task experience gives stable gains that do not grow with more samples, and agents fail to extract useful signal from extra experience. Code and data are published by the PKU Value Lab.
- EmbodiedCLBench is an embodied continual-learning benchmark for compositional generalization.
- Experience from basic tasks improves hard tasks but does not scale with more samples.
- Zero-shot performance on advanced tasks stays weak across tested agent harnesses.
- Code and dataset are released by the PKU Value Lab.
- Organizations
- Peking University
From the original article
1 min read at huggingface.co
Continual learning requires models to accumulate past experience and transfer it to new tasks. Despite substantial research on continual learning, existing work remains largely confined to non-embodied settings, without extending to embodied environments.
By Jie Huang, Yanan Chen, Ruixun Liu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.