PyoSignal Logo
PyoSignal
Back to Research

RynnWorld-4D: 4D Embodied World Models for Robotic Manipulation

Paper ID: 2607.06559 β€’ 72 Upvotes
Robotics World Models Diffusion Models 4D Vision Embodied AI Vision Video
RynnWorld-4D: 4D Embodied World Models for Robotic Manipulation

πŸ“ 핡심 μš”μ•½

RGB-D와 κ΄‘ν•™ 흐름을 κ²°ν•©ν•œ 4D μ›”λ“œ λͺ¨λΈμ„ 톡해 λ‘œλ΄‡μ˜ μ •λ°€ν•œ 양손 μ‘°μž‘ μ„±λŠ₯을 κ·ΉλŒ€ν™”ν•œ 연ꡬ

πŸ”‘ μ£Όμš” λ‚΄μš© (Key Points)

  • RGB-D와 Optical Flowλ₯Ό κ²°ν•©ν•˜μ—¬ κΈ°ν•˜ν•™μ  ꡬ쑰와 μ›€μ§μž„μ„ λ™μ‹œμ— ν¬μ°©ν•˜λŠ” 4D ν‘œν˜„ 방식 μ œμ•ˆ
  • 단일 ν™•μ‚° ν”„λ‘œμ„ΈμŠ€ λ‚΄μ—μ„œ μ‹œκ°μ  μ™Έν˜•, 깊이, μ›€μ§μž„μ„ μΌκ΄€λ˜κ²Œ μƒμ„±ν•˜λŠ” Tri-branch μ•„ν‚€ν…μ²˜ 섀계
  • 2.5μ–΅ ν”„λ ˆμž„ μ΄μƒμ˜ λŒ€κ·œλͺ¨ 데이터셋(Rynn4DDataset 1.0) ꡬ좕 및 μ—­μ—­ν•™ ν—€λ“œλ₯Ό ν†΅ν•œ 효율적 μ •μ±… ν•™μŠ΅

πŸ’‘ 싀무적 κ°€μΉ˜ (Relevance)

λ‹¨μˆœ 2D λΉ„λ””μ˜€ μ˜ˆμΈ‘μ„ λ„˜μ–΄ 물리적 ꡬ쑰가 반영된 4D 데이터λ₯Ό μƒμ„±ν•¨μœΌλ‘œμ¨, λ‘œλ΄‡μ˜ 말단 μž₯치(End-effector) μ œμ–΄μ— μ¦‰μ‹œ ν™œμš© κ°€λŠ₯ν•œ ν‘œν˜„ 곡간을 μ œκ³΅ν•©λ‹ˆλ‹€.

βœ… μΆ”μ²œ μ•‘μ…˜ (Actionable Items)

  • 제곡된 Rynn4DDataset의 데이터 ν’ˆμ§ˆκ³Ό Pseudo-label의 신뒰도 뢄석
  • RynnWorld-4D-Policy의 μ—­μ—­ν•™ ν—€λ“œλ₯Ό κΈ°μ‘΄ λ‘œλ΄‡ μ œμ–΄ μ•Œκ³ λ¦¬μ¦˜κ³Ό κ²°ν•©ν•˜λŠ” μ‹€ν—˜
  • λ‹€μ–‘ν•œ μ‘°μž‘ ν™˜κ²½μ—μ„œμ˜ 4D 생성 λͺ¨λΈμ˜ μ‹œκ³΅κ°„μ  일관성 검증