README.md

---
license: apache-2.0
---

<h1 align="center">EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents</h1>

<p align="center"><em>The first benchmark in embodied simulation for evaluating<br>
the continual learning capability of self-evolving agents</em></p>

<p align="center">
  <a href="https://arxiv.org/abs/2610.13021"><img src="https://img.shields.io/badge/arXiv-2610.13021-b31b1b" alt="arXiv: 2610.13021"></a>
  <a href="https://pku-value-lab.github.io/EmbodiedCLBench-homepage/"><img src="https://img.shields.io/badge/Project-Page-blue" alt="Project Page"></a>
  <a href="https://huggingface.co/datasets/PKU-VaLuE-Lab/EmbodiedCLBench"><img src="https://img.shields.io/badge/Hugging%20Face-Dataset-yellow" alt="Hugging Face Dataset"></a>
  <a href="https://github.com/PKU-VaLuE-Lab/EmbodiedCLBench"><img src="https://img.shields.io/badge/GitHub-Code-black" alt="GitHub Code"></a>
</p>

<p align="center">
  <a href="https://jie-1203.github.io/">Jie Huang</a><sup>★,1,2</sup>&emsp;
  <a href="https://github.com/YananChen8">Yanan Chen</a><sup>★,1</sup>&emsp;
  <a href="https://liuruixun.github.io/">Ruixun Liu</a><sup>★,1</sup>&emsp;
  <a href="https://hkc20.github.io/">Kaichen He</a><sup>1</sup>
  <br>
  <a href="https://github.com/hexinyi2101">Xinyi He</a><sup>1</sup>&emsp;
  <a href="https://ruihuangai.github.io/">Rui Huang</a><sup>3</sup>&emsp;
  <a href="https://zilongzheng.github.io/">Zilong Zheng</a><sup>2</sup>&emsp;
  <a href="https://www.zlz.link/">Zhenliang Zhang</a><sup>2</sup>&emsp;
  <a href="https://yiwuzhong.notion.site/homepage">Yiwu Zhong</a><sup>1,2,†</sup>
</p>

<p align="center">
  <sup>1</sup>School of Intelligence Science and Technology, Peking University<br>
  <sup>2</sup>State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing, China<br>
  <sup>3</sup>University of Hong Kong
</p>

<p align="center">★ Equal contribution.&emsp;† Corresponding author.</p>

## Core design

<p align="center">
  <img src="https://pku-value-lab.github.io/EmbodiedCLBench-homepage/static/images/paper/intro.png" alt="EmbodiedCLBench overview showing compositional generalization from basic to advanced tasks" width="100%">
</p>

We operationalize continual learning through compositional generalization: an agent first completes simpler tasks to accumulate experience, then is tested on whether it can recombine the acquired skills to solve more complex, unseen tasks.

## Abstract

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.

To address this gap, we introduce EmbodiedCLBench, the first benchmark in embodied simulation for evaluating the continual learning capability of self-evolving agents. With compositional generalization as the core evaluation principle, our benchmark evaluates whether agents can first learn from basic tasks and then recombine the acquired skills to solve complex, unseen advanced tasks. Based on EmbodiedCLBench, we conduct extensive experiments across multiple agent harnesses and learning methods, revealing consistent limitations and distinctive behaviors.

Our results show that while advanced tasks can be hardly solved under zero-shot setting, the experience from basic tasks enables consistent improvement. Such benefit from the experience remains stable even if the task complexity increases. However, this improvement does not scale accordingly with the number of learning samples, and current agents cannot extract meaningful information from additional experience. Collectively, our findings highlight continual learning as a fundamental yet underexplored capability, and our benchmark offers valuable resources for designing effective self-evolving agents in embodied environments.

## Benchmark Construction

<p align="center">
  <img src="https://pku-value-lab.github.io/EmbodiedCLBench-homepage/static/images/paper/construction.png" alt="Overview of the EmbodiedCLBench construction process" width="100%"><br>
  Overview of the construction process
</p>

<br>

<p align="center">
  <img src="https://pku-value-lab.github.io/EmbodiedCLBench-homepage/static/images/paper/statistics.png" alt="EmbodiedCLBench benchmark statistics" width="100%"><br>
  Benchmark statistics
</p>

## Main Results

<p align="center">
  <img src="https://pku-value-lab.github.io/EmbodiedCLBench-homepage/static/images/paper/main-results.png" alt="Main results: success rates on basic and advanced tasks across harnesses, models, and learning settings" width="100%">
</p>

(1) **Zero shot:** the agent attempts tasks directly without prior experience. (2) **In-Context Learning (ICL):** the agent first completes two related basic tasks, retaining the full interaction history in context for the advanced task. (3) **Skill Learning:** the same learning phase as ICL, but the agent distills its experience into a concise skill summary, which replaces the raw trajectories at test time.

## Citation

```bibtex
@misc{huang2026embodiedclbench,
  title         = {EmbodiedCLBench: Evaluating Continual Learning
                   for Self-Evolving Embodied Agents},
  author        = {Jie Huang and Yanan Chen and Ruixun Liu and
                   Kaichen He and Xinyi He and Rui Huang and
                   Zilong Zheng and Zhenliang Zhang and Yiwu Zhong},
  year          = {2026},
  eprint        = {2610.13021},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2610.13021},
}
```