dingxingdi Skill: landmark-grounded-semantic-exploration
1. Capability Definition & Real Case
- Professional Definition: This capability measures whether an embodied agent can use landmarks, chained spatial cues, and large-scale exploration to find the exact region or object needed for an answer in an open environment. It requires cross-scale navigation: long-range movement to the relevant area and local inspection to collect the final evidence.
- Dimension Hierarchy: Goal-Directed Exploration->Question-Conditioned Exploration->landmark-grounded-semantic-exploration
Real Case
[Case 1]
- Initial Environment: A UAV-style embodied agent starts above a city block. Directly south is a large building, and east of that building is a shop with a yellow signboard. Two different cars are visible in the wider area.
- Real Question: There is a building to the south of you. To the east of the building is a shop with a yellow signboard. Please tell me what color is the car parked in front of the shop.
- Real Trajectory: The agent first verifies the southern building, travels to its east side, locates the yellow-sign shop, narrows attention to the car in the frontal parking position, and then inspects color.
- Real Answer: Red.
- Why this demonstrates the capability: The question is intentionally ambiguous without the landmark chain. The agent must use landmark grounding to filter the candidate area before doing local perception. This matches the intended long-horizon city exploration pattern.
[Case 2]
- Initial Environment: The agent begins at a mid-altitude viewpoint over a commercial street with several storefronts, parked vehicles, and signs. A brown awning shop is visible only after the agent shifts laterally along the block.
- Real Question: Can I get coffee from the shop with the brown awning?
- Real Trajectory: The agent explores the storefront sequence, identifies the shop with the brown awning, inspects signage and exterior cues, and then answers based on the shop’s function.
- Real Answer: Yes.
- Why this demonstrates the capability: The agent cannot answer from a single isolated storefront crop. It must navigate a large scene, bind the right landmark, and then interpret semantic evidence. This combines exploration with grounded world knowledge.
[Case 3]
- Initial Environment: The agent starts above a parking district where a named store, a lot, and several lanes are visible but small because of distance. The lot contains multiple vehicles and partial occlusion from a street tree.
- Real Question: How many cars are parked in the parking lot?
- Real Trajectory: The agent first identifies the correct lot by landmark cues, descends or moves closer for a finer view, and then counts only the vehicles inside the lot boundary.
- Real Answer: Eight.
- Why this demonstrates the capability: The challenge is not just counting. The agent must first determine which region counts as the target lot and then refine its observation scale to avoid missing small or distant cars. This is the signature pattern of landmark-grounded exploration in city environments.
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:
Phase 1: Environment Exploration Read the exploration guidelines to discover raw knowledge seeds:
references/EXPLORATION.mdPhase 2: Trajectory Selection Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
references/SELECTION.mdPhase 3: Data Synthesis Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
references/SYNTHESIS.md