EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents

EmbodiedCLBench overview showing compositional generalization from basic to advanced tasks — **Core design:** 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

Overview of the EmbodiedCLBench construction process — Overview of the construction process

EmbodiedCLBench benchmark statistics — Benchmark statistics

Results and Findings

Main results.

Success Rate (%)		Basic Task	Advanced Task
Harness	Model	Zero-shot	Zero-shot	In-Context Learning	Skill Learning
Hermes	Qwen3.8-Flash	58.9	8.0	44.0	33.0
Hermes	GPT5.6-Terra	53.0	11.0	16.0	34.0
OpenClaw	Qwen3.8-Flash	50.6	5.0	35.0	38.0
OpenClaw	GPT5.6-Terra	47.6	12.0	39.0	22.0
Codex	Qwen3.8-Flash	52.4	4.0	34.0	33.0
Codex	GPT5.6-Terra	54.8	11.0	52.0	26.0
Claude Code	Qwen3.8-Flash	42.9	3.0	27.0	31.0
Claude Code	GPT5.6-Terra	51.8	9.0	31.0	25.0

(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.

Effect of Compositional Complexity

Does more complex composition make continual learning harder?

Scaling with the Number of Learning Samples

Does more experience lead to a higher success rate on advanced tasks?

BibTeX

@misc{huang2026embodiedclbench,
  title   = {EmbodiedCLBench: Evaluating Continual Learning
             for Self-Evolving Embodied Agents},
  author  = {Huang, Jie and Chen, Yanan and Liu, Ruixun and
             He, Kaichen and He, Xinyi and Huang, Rui and
             Zheng, Zilong and Zhang, Zhenliang and Zhong, Yiwu},
  year    = {2026},
  url     = {https://pku-value-lab.github.io/EmbodiedCLBench-homepage/},
}