spatial-relation-reasoning

# Skill: spatial-relation-reasoning ## 1. Capability Definition & Real Case * **Professional Definition**: This capability measures whether an embodied agent can infer discrete or descriptive spatial relationships among objects, landmarks, and egocentric viewpoints. It requires consistent frame selection and relational reasoning, especially when the same scene can support different answers under different viewpoints. * **Dimension Hierarchy**: Perceptual World Modeling->Static Spatial Grounding->spatial-relation-reasoning ### Real Case **[Case 1]** * **Initial Environment**: The agent stands on a city street facing a building. To the east of that building is a shop with a yellow signboard, and a red car is parked directly in front of the shop while another car is parked farther down the road. * **Real Question**: What color is the car parked in front of the shop with the yellow signboard? * **Real Trajectory**: The agent first localizes the building used as a landmark, then moves to its east side, then identifies the shop with the yellow signboard, and finally inspects the car occupying the in-front-of relation. * **Real Answer**: Red. * **Why this demonstrates the capability**: The scene contains multiple cars, so the target is only identifiable through chained spatial relations. The agent must preserve the relation 'east of building' and then 'in front of shop' to avoid answering about the wrong vehicle. This is a direct test of relation-based disambiguation. --- **[Case 2]** * **Initial Environment**: A living room scene contains a sofa, a coffee table, a lamp behind the sofa, and a chair to the left of the coffee table from the agent’s current viewpoint. * **Real Question**: Is the chair on the left or right of the table? * **Real Answer**: Left. * **Why this demonstrates the capability**: The answer depends on the agent’s current frame of reference and not merely object identity. If the agent silently flips the viewpoint or uses an allocentric frame inconsistently, it will fail. The task therefore probes viewpoint-consistent relational reasoning. --- **[Case 3]** * **Initial Environment**: A service robot observes a doorway with a clock above it, a portrait on the wall, and a coat hanger on the floor near the left side of the scene. The robot is asked for navigation advice rather than a label. * **Real Question**: Where should I go if I want to find the coat hanger? * **Real Answer**: Go toward the doorway and look on the floor just left of the portrait. * **Why this demonstrates the capability**: The agent must transform raw spatial relations into a usable directional description. It has to preserve above, left-of, and near relations while producing an action-relevant answer. That makes the problem relational and embodied rather than purely descriptive. ## 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`