SEAR：自进化智能体机器人技术
SEAR: Self-Evolving Agentic Robotics
AI总结：
本研究提出SEAR框架与基准，探究智能体机器人的学习本质与泛化性，发现其学习曲线受预算与任务难度影响，控制代码可跨场景迁移，不同模型进化方式有差异，且纯文本模型可与多模态模型性能相当。
摘要：
前沿语言模型智能体现在可以自行改进机器人：它们编写控制代码、测试代码，并将有效的部分作为技能和记忆保留下来，用于下一个任务。我们推出SEAR（Self-Evolving Agentic Robotics，自进化智能体机器人技术），这是一个围绕两个核心问题构建的框架和基准：智能体是真的在学习，还是只是尝试得更频繁？它学到的东西是否仍能在新任务、新模拟器和真实机器人上发挥作用？该框架让智能体以三种模式运行：编写成为策略的代码（代码即策略）、在控制循环中实时行动（智能体即策略），或作为微调视觉-语言-动作模型的研究者（自动研究）。该基准添加了控制项，将保留的经验与重复尝试、熟悉任务、特权模拟器状态、更大预算和运气分离开来。我们在7个前沿语言模型、330个模拟任务和5个物理任务，以及超过4年的智能体运行时间的实验中，得出三个关键发现：第一，机器人智能体的学习曲线既取决于总体时间预算，也取决于任务相对于模型能力的难度；更难的任务值得更长的会话时间，而对于较简单的任务，重复进行短时长尝试的效果同样好。第二，控制代码和共享例程可跨任务和模拟器迁移，而特定任务的解决方案则无法迁移；在真实机器人上，我们观察到智能体设计的实践与人类设计的新场景之间存在性能差距。第三，不同模型的进化方式效率不同：一些模型在实时行动时表现最佳，另一些则在先编写代码时表现最佳；此外，纯文本模型与可比的多模态模型表现相当：感知有助于错误评估，但智能体推理和工具使用可能更重要。我们发布该基准和代码库，以支持未来的研究。
英文摘要：
Frontier language-model agents can now improve robots on their own: they write control code, test it, and keep what works as skills and memories for the next task. We introduce SEAR (Self-Evolving Agentic Robotics), a framework and benchmark built around two questions. Is the agent really learning, or just trying more often? And does what it learns still help on new tasks, in a new simulator, and on a real robot? The framework runs an agent in three modes: it writes code that becomes the policy (code-as-policy), acts live in the control loop (agent-as-policy), or works as a researcher that fine-tunes a vision-language-action model (auto-research). The benchmark adds controls that separate retained experience from repeated attempts, familiar tasks, privileged simulator state, larger budgets, and luck. Across seven frontier language models, 330 simulated and five physical tasks, and over four years of agent runtime, we highlight three findings. First, the learning curve of a robotic agent depends both on the overall time budget and the task difficulty relative to model capability. Harder tasks make longer sessions worthwhile; on easier tasks, re-rolling short attempts does just as well. Second, control code and shared routines transfer across tasks and simulators, while task-specific solutions do not. On a real robot, we observe a performance gap between agent-designed practice and novel human-designed scenes. Third, different models evolve more effectively in different ways. Some do best acting live, others writing code first. And a text-only model keeps pace with a comparable multimodal one: perception helps with error assessment, but agentic reasoning and tool use can matter more. We release the benchmark and codebase to support future research.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
