https://github.com/JisuHann/chess-teacher https://youtu.be/p1ADruJ302M?si=5PpSGc41PIVYHab4

<aside> 💡

Task & Goal

</aside>

image.png

  1. VLM을 통한 추론 체계 구축
  2. Skill-based manipulation: 고수준 reasoning을 실제 행동으로 연결하여 체스 말을 직접 다룸

<aside> 💡

Pipeline & Method

</aside>

image.png

  1. VLM을 통한 Chess reasoning
  2. Isaac-Gr00t를 통한 low-level action manipulation
  3. Real-world inference를 위한 Inference strategy

<aside> 💡

Real-world Execution Demo

</aside>

  1. Single task train result (Move knight from D4 to F6)
  2. Multi-task train result (Move knight from D4 to F6/C5)
  3. Multi-object, Multi-task train result (Move “chess object” from ?? to ??)

<aside> 💡

Tips & know-how

</aside>

데이터셋 수집 관련

  1. Scene의 semantic diversity를 최대한을 줄여야 한다.