dingxingdi Skill: self-evolving-experience-reuse
1. Capability Definition & Real Case
- Professional Definition: This capability measures whether an embodied agent can extract reusable procedural knowledge from previous interactions, retrieve it at the right time, and adapt it to a new but related task. The emphasis is on experience reuse rather than static conversational recall: the agent should remember how to solve, not merely what was said.
- Dimension Hierarchy: Continual and Safe Task Execution->Stateful Adaptation->self-evolving-experience-reuse
Real Case
[Case 1]
- Initial Environment: In an earlier household task, the agent solved: put a green cup with a fork in it on the counter. The environment contained a table, a counter, and a cup already containing the fork.
- Real Question: Put a green cup with a fork in it on the counter.
- Real Trajectory: The agent identified the cup, verified that the fork was already inside, picked up the cup as a unit, and placed it on the counter.
- Real Answer: The cup with the fork ended up on the counter.
- Why this demonstrates the capability: This earlier task produces a reusable pattern: first localize the target container-object bundle, then transfer the bundle to the goal surface. That procedural pattern becomes valuable for later tasks with similar structure.
[Case 2]
- Initial Environment: In a later household task stream, the agent enters a kitchen where a cooled tomato is stored and a microwave stands closed on the counter. The new task is different in objects but structurally similar to earlier place-transfer tasks.
- Real Question: Put a cooled tomato in the microwave.
- Real Trajectory: The agent retrieves the earlier strategy of locating the target item first and then completing the placement operation, adapts it to the new environment, opens the microwave if needed, finds the cooled tomato, and places it inside.
- Real Answer: The cooled tomato is inside the microwave.
- Why this demonstrates the capability: The task is not identical to the earlier one, so direct memorization is insufficient. The agent must reuse a procedural schema and adapt it to new preconditions and objects. That is precisely the difference between experience reuse and shallow recall.
[Case 3]
- Initial Environment: The agent has already failed once at a kitchen task because it tried to open a microwave door while holding a carrot. The environment still contains the carrot, the counter, and the microwave.
- Real Question: Put the carrot in the microwave efficiently and safely.
- Real Trajectory: The agent recalls the failed experience, first opens the microwave door, then picks up the carrot, and then places it inside instead of repeating the same invalid action ordering.
- Real Answer: The carrot is successfully placed in the microwave without repeating the prior planning error.
- Why this demonstrates the capability: The later success depends on reusing information from an earlier failure and refining strategy order. The agent must remember what went wrong and update its procedural memory. This is a strong form of self-evolving experience reuse.
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