PKU Value Lab Unveils EmbodiedCLBench

Researchers from the PKU VALUE Lab have released EmbodiedCLBench, a novel simulation benchmark created to evaluate continual learning capabilities in self-evolving embodied agents. Built around the core principle of compositional generalization, the benchmark tests whether agents can acquire foundational skills from basic tasks and recombine those abilities to complete complex, previously unseen objectives in embodied environments. The project's code and datasets have been made publicly available.

Extensive testing across various agent harnesses and learning methods revealed consistent behavioral patterns across current model designs. Under standard zero-shot conditions, agents struggle significantly, showing that advanced embodied tasks remain virtually unsolvable without prior learning. However, exposing models to basic foundational tasks enables consistent performance gains on downstream challenges, with this positive experience transfer remaining stable even as task complexity escalates.

Despite these clear benefits from initial experience, the benchmark highlighted a major bottleneck in how current models process ongoing data. The researchers discovered that performance improvements do not scale effectively with an increase in the number of learning samples. Present agent architectures fail to extract additional meaningful information from supplementary experience, causing skill acquisition to plateau regardless of extended training exposure.

For AI practitioners and robotics developers, EmbodiedCLBench shifts the focus from static, single-task evaluations toward long-term adaptable intelligence. By identifying how agents store and reconfigure functional skills over time, the benchmark offers crucial diagnostic tools for building self-evolving autonomous systems. Addressing the failure of models to benefit from sample scaling will be key to unlocking agents that can continuously adapt to complex real-world environments.