2025.07.30 - #44 - Feed-forward 3D recon survey, ThinkAct, ROMAN, UA-MPC, TurboClique

[changh95]

Interesting papers

Advances in Feed-Forward 3D Reconstruction and View Synthesis: A Survey

• https://www.arxiv.org/abs/2507.14501
• NeRF, Pointmap, 3D Gaussian Splatting, Mesh / Occupancy / SDF, 3D representation-free 모델 기법들의 계보 설명

ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning

• https://jasper0314-huang.github.io/thinkact-vla/
• 기존의 VLA 모델은 단순히 Input->output 맵핑 형태임
• ThinkAct는 Multi-modal reasoning LLM을 사용해서 trajectory/action에 대한 latent variable을 생성하는데, 이걸 downstream에 있는 action model에 넣으면 trajectory가 나온다.
• 여기서 Multi-modal reasoning LLM이 latent variable을 잘 생성해줘야 좋은 trajectory가 나올텐데, 이건 visual 데이터로 평가하는 goal reward + trajectory reward를 이용한 강화학습 (GRPO)를 이용해서 LLM을 학습한다.

ROMAN

• https://github.com/mit-acl/roman
• Open-set detection object들을 이용해서 localization

[U-AMC]

Interesting papers

UA-MPC: Uncertainty-Aware Model Predictive Control for Motorized LiDAR Odometry

https://github.com/kafeiyin00/UA-MPC

TurboClique

https://github.com/Laka-3DV/TurboReg

TurboClique for Robust and Efficient Point Cloud Registration

Registration beyond Points: General Affine Subspace

Alignment via Geodesic Distance on Grassmann Manifold

[changh95]

Lie group 강의: https://drive.google.com/file/d/1fZNbBuGz0mDN00ZAMx_FNSRGUfHv1neF/view?usp=sharing

Lie algebra 강의: https://drive.google.com/file/d/1g-tUQLwnO0QEsRuoxAmCeODmCoEmv0rl/view?usp=sharing

[changh95]

https://link.springer.com/book/10.1007/978-3-319-13467-3

https://media.springernature.com/w316/springer-static/cover-hires/book/978-3-319-13467-3?as=webp

[U-AMC]

https://arxiv.org/abs/2409.11935