面向长程具身操作的神经符号过程推理

Translation Process Summary

Alright, let's break this down. The task is to translate this research abstract into Simplified Chinese, maintaining the original paragraph structure and avoiding any extraneous commentary. My process is pretty straightforward: I start by dissecting each sentence, identifying key terms and their most accurate Chinese equivalents. "Vision-language-action (VLA) models" is a common one, so I'll render it as "视觉-语言-动作（VLA）模型". I need to be precise, especially with terms like "grounding," which I'll translate here as "具身定位" or "感知定位."

I work sentence by sentence. For example: "VLA models can execute short manipulation skills, but remain brittle in long-horizon procedures..." becomes "视觉-语言-动作（VLA）模型能够执行简短的操作技能，但在面对需要持久任务状态、依赖感知推理、条件决策和可靠具身定位的长程流程时，依然表现脆弱。" I'm paying attention to the nuances of each term and how they fit within the context of robotics and VLA research. "Neuro-symbolic framework" translates neatly to "神经符号框架."

I ensure the translation is fluent and retains the original meaning. For instance, "Task graphs encode action dependencies, valid transitions, and branch conditions..." requires a clear translation: "任务图对动作依赖关系、有效状态转移及分支条件进行编码..." I maintain the technical accuracy and, at the same time, aim for natural phrasing in Chinese. I then refine the translation, ensuring smooth flow and coherence. I also double-check the translations of phrases like "destination grounding" to be as precise as possible – I settle on "目标位置定位" in this instance.

Finally, I present the translation, adhering strictly to the prompt's instructions: no commentary, just the finished Chinese text.

视觉-语言-动作（VLA）模型能够执行简短的操作技能，但在需要持久任务状态、依赖感知推理、条件决策和可靠定位的长程任务流程中仍然表现脆弱。我们研究了一种神经符号框架，该框架将习得的 VLA 控制与显式任务图以及多模态流程记忆相结合。任务图对动作依赖关系、有效状态转移和分支条件进行编码，而记忆模块则维护当前步骤、已完成动作、文本上下文以及与任务相关的视觉证据。这些结构协同指导物体选择、目标位置定位、子目标调度以及对预期状态转移的验证。人类示范则通过视线或显著性线索提供额外的时空引导。为了单独分析其对策略学习的影响，我们的初步研究绕过了跨视角视线迁移，直接在机器人视角的遥操作视频中对伪视线进行标注。由此产生的引导信息被应用于 VLA 的微调和推理阶段。我们研究了两个需要有序执行、基于视觉定位的决策以及条件分支的长程操作领域——工作区清理和手术器械操作。我们评估了正确物体与目标位置的选择、子任务完成情况、任务进度、步骤顺序一致性、整项任务成功率以及流程或执行错误。本工作将结构化符号推理与源自示范的视觉引导确立为实现可靠长程 VLA 操作的互补机制。