Self-Evolving Embodied Agents via Skill-Harness Evolution

通过技能与工具链演化实现具身智能体的自我进化

Embodied agents are increasingly built as systems around foundation models, where performance depends not only on model weights but also on the skills, context, action interfaces, and execution harness surrounding the model.

具身智能体正越来越多地围绕基础模型构建为系统，其性能不仅取决于模型权重，还取决于围绕模型构建的技能、上下文、动作接口以及执行工具链（execution harness）。

While supervised fine-tuning and reinforcement learning can adapt agents to new environments, they require additional data, rewards, and training runs; meanwhile, many train-free code-centric approaches rely on programmable robot APIs that may be unavailable in fixed-interface settings.

虽然监督微调和强化学习可以使智能体适应新环境，但它们需要额外的数据、奖励和训练过程；与此同时，许多无需训练的以代码为中心的方法依赖于可编程的机器人 API，而这些 API 在固定接口的环境中可能无法使用。

We propose SHAPER, a self-evolving framework for train-free embodied adaptation that keeps model parameters frozen and improves the non-parametric agent system by evolving reusable skills and a context-code harness through target-environment rollouts.

我们提出了 SHAPER，这是一个用于无需训练的具身适应的自我进化框架。它保持模型参数冻结，并通过在目标环境中进行试运行（rollouts），演化可重用的技能和上下文代码工具链，从而改进非参数化的智能体系统。

In SHAPER, the same frozen model can serve as both planner and optimizer, refining its external skills and context-code harness without parameter updates.

在 SHAPER 中，同一个冻结模型既可以充当规划器，也可以充当优化器，在无需更新参数的情况下优化其外部技能和上下文代码工具链。

We evaluate SHAPER on VLABench and ESI-Bench, covering embodied agents with different low-level action interfaces, and compare against pure execution, supervised fine-tuning, and test-time-scaling baselines such as verifier-free selection and voting.

我们在 VLABench 和 ESI-Bench 上评估了 SHAPER，涵盖了具有不同底层动作接口的具身智能体，并将其与纯执行、监督微调以及测试时扩展基准（如无验证器选择和投票）进行了比较。

Our results suggest that skill-and-harness optimization is a practical route to self-evolving embodied agents when model training is expensive, unavailable, or undesirable.

我们的研究结果表明，当模型训练成本高昂、无法进行或不理想时，技能与工具链优化是实现具身智能体自我进化的一种实用途径。