返回 Skill 列表
extension
分类: 内容与媒体无需 API Key

scienceworld-target-locator

这项技能根据领域知识和环境线索确定目标对象最可能的位置。当代理需要找到特定物品(如动物)但该物品不在当前房间时触发此技能。它会分析环境描述并建议传送到哪个房间进行进一步调查。

person作者: jakexiaohubgithub

Skill: Target Locator

Trigger: When the agent needs to find a specific object (e.g., an animal, a tool, a chemical) and a preliminary look around in the current room does not reveal it.

Core Logic

  1. Analyze the Target: Classify the target object based on its type (e.g., animal, tool, container, chemical).
  2. Consult Domain Knowledge: Use the bundled reference (target_location_heuristics.md) to map the object type to the most probable room(s) in the ScienceWorld environment.
  3. Evaluate Current Context: Briefly review the recent look around observation. If doors to high-probability rooms are visible and open, prioritize them.
  4. Output Decision: Output a single, clear teleport to LOC action command, choosing the most promising location to search next.

Primary Instruction: Do not overthink. Rely on the heuristics. The goal is to make an efficient, educated guess to continue the search.

Example Flow (From Trajectory)

  • Task: "find a(n) animal."
  • Trigger: look around in hallway shows no animal.
  • Analysis: Target type = animal.
  • Heuristic Check: Animals are most commonly found outside.
  • Context Check: A door to outside is not listed in the hallway observation.
  • Decision: The highest-probability room outside is accessible via teleport.
  • Output: teleport to outside