Paper review · October 13, 2026
EmbodiedCLBench: Evaluating Continual Learning in Self-Evolving Embodied Agents
EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents
The authors build an embodied simulation benchmark that measures whether agents can combine experience from basic tasks to solve unseen advanced tasks, and show that experience improves performance but gains do not scale with the number of samples.
Problem
Continual learning research and benchmarks have mostly stayed in digital and symbolic environments. There was no dedicated benchmark for evaluating whether agents build up experience and transfer it to new tasks in embodied environments with physical constraints and spatial complexity.
Method
Compositional generalization is the evaluation principle. Based on the TongSim household simulation, the authors built a three-level task hierarchy: atomic actions → basic tasks (2 atomic actions, at least 6 steps) → advanced tasks (combining several basic-task categories, 4 atomic actions, at least 12 steps). For reproducibility, each task is built as an offline state-transition graph with first-person images and candidate actions. Candidate actions are labeled optimal, suboptimal or infeasible. In the learning phase, the agent performs two related basic tasks and then solves an advanced task. There are three evaluation settings: Zero-shot, In-Context Learning (Trajectories kept in context as-is) and Skill Learning (only experience summaries provided).
Results
- The benchmark consists of 268 tasks in total, including 168 basic tasks and 100 advanced tasks.
- Across combinations of 4 harnesses (OpenClaw, Claude Code, Codex, Hermes) and 2 models, Zero-shot success rates were about 50% on basic tasks and about 10% on advanced tasks.
- With experience, advanced-task success rates generally rose above 30%. For example, Codex+GPT5.6-Terra went from 11.0% Zero-shot to 52.0% with ICL.
- Neither ICL nor Skill Learning was consistently better. Increasing the number of atomic actions to 3–6 kept the improvement over Zero-shot roughly the same.
- Even a single basic task raised performance substantially, but adding more training samples gave no consistent further gain.
Limitations
Evaluation is offline and multiple-choice: agents pick candidate actions on precomputed state graphs rather than performing continuous control. The simulator, TongSim, is closed, and only indoor household scenes are covered. Task descriptions and candidate actions were generated by an LLM and then verified by humans. Success rate is the only metric.
Why it matters
It is the first systematic measurement of whether embodied agents can learn from experience and transfer it to harder tasks. It reveals that current agents fail to extract additional information as experience grows, which points to open problems in designing self-evolving agents.
Key terms
- Continual Learning
- A model's ability to keep accumulating past experience and use it to solve new tasks.
- Compositional Generalization
- The ability to combine separately learned basic elements in new ways to solve unseen complex problems.
- Agent Harness
- An agent execution framework, such as OpenClaw, Claude Code or Codex, that wraps a model and provides tool calls, memory and an execution loop.
Original
Authors: Jie Huang, Yanan Chen, Ruixun Liu, Kaichen He, Xinyi He, Rui Huang, Zilong Zheng, Zhenliang Zhang, et al. · Posted 2026-10-09 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.