exploration-aware-evidence-collection

# Skill: exploration-aware-evidence-collection ## 1. Capability Definition & Real Case * **Professional Definition**: This capability measures whether an embodied agent can plan and execute an exploration trajectory that collects sufficient task-relevant evidence before answering an open-ended question. It emphasizes grounding, evidence sufficiency, and the alignment between what the agent observed and what it claims in the final answer. * **Dimension Hierarchy**: Goal-Directed Exploration->Question-Conditioned Exploration->exploration-aware-evidence-collection ### Real Case **[Case 1]** * **Initial Environment**: The agent starts in a home corridor outside the kitchen with only a partial view of cabinets and the sink area. The faucet itself is not visible from the start position. * **Real Question**: Did I leave the faucet running in the kitchen? * **Real Trajectory**: The agent walks from the corridor into the kitchen, turns toward the sink, inspects the faucet directly, and stops once the water state is visible. * **Real Answer**: No, the faucet is turned off. * **Why this demonstrates the capability**: The answer cannot be grounded from the starting observation. The agent must move to a location where the target state becomes directly observable. This tests exploration quality and evidence gathering rather than prior knowledge. --- **[Case 2]** * **Initial Environment**: The embodied agent begins near the entryway of an apartment. The living room and TV corner are not fully visible, and the floor lamp is hidden behind a partial wall until the agent enters deeper into the room. * **Real Question**: Where is the floor lamp in the living room? * **Real Trajectory**: The agent explores the living room, checks the TV side of the room, identifies the lamp, and records a grounded spatial answer tied to a visible landmark. * **Real Answer**: It is in the corner of the living room, next to the TV. * **Why this demonstrates the capability**: The correct answer depends on reaching a vantage point with clear evidence. A generic guess like 'near the couch' would be plausible but ungrounded. The capability is therefore about collecting the right visual proof before responding. --- **[Case 3]** * **Initial Environment**: The agent starts in a laundry-adjacent hallway and has not yet observed the utility area. Several rooms are reachable from the initial location, including a bathroom and a bedroom. * **Real Question**: How many washing machines do I have? * **Real Trajectory**: The agent first identifies the laundry region, enters it, inspects the appliance area, and counts distinct washing machine units before answering. * **Real Answer**: Only one. * **Why this demonstrates the capability**: The task is not answerable from commonsense priors because the home could plausibly contain zero, one, or multiple machines. The agent must collect counting evidence from a relevant region. This makes the exploration path itself part of the competence being tested. ## Pipeline Execution Instructions To synthesize data for this capability, you must strictly follow a 3-phase pipeline. **Do not hallucinate steps.** Read the corresponding reference file for each phase sequentially: 1. **Phase 1: Environment Exploration** Read the exploration guidelines to discover raw knowledge seeds: `references/EXPLORATION.md` 2. **Phase 2: Trajectory Selection** Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: `references/SELECTION.md` 3. **Phase 3: Data Synthesis** Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: `references/SYNTHESIS.md`