Research

VGGRPO: 4D Geometry Reward for Video Diffusion

Video diffusion model의 3D consistency를 4D geometry reward + GRPO로 강화하는 방법론 분석 및 로봇 도메인 적응 전략
6 References
4 Open Questions
4 Technical Challenges

Research Question

Video Diffusion + 4D Geometry Reward + GRPO
Video diffusion model의 3D consistency를 4D geometry reward로 강화할 수 있는가? LLM 영역의 GRPO가 diffusion model에도 적용 가능한가? 로봇 도메인(dual-view, ego+wrist)에서 이 접근이 유효한가?
4D geometry reward GRPO for diffusion 3D consistency video generation robotics adaptation

Source Paper: VGGRPO

Towards World-Consistent Video Generation with 4D Latent Reward
Authors: Zhaochong An, Orest Kupyn, et al. (Google, U Copenhagen, Oxford)
arXiv: 2603.26599
Core idea: Video diffusion으로 multi-view video 생성 → 4D reconstruction으로 geometry 평가 → reward signal로 GRPO 최적화

Architecture Pipeline

Video Diffusion (SV4D) Generated multi-view video 4D Reconstruction (DUSt3R/MASt3R) 4D Gaussians Reward Computation: r = α·SSIM_novel + β·(1 - depth_error) + γ·flow_consistency GRPO: policy gradient with KL constraint updated diffusion weights

Key Claims

Our Adaptation for Robotics

Paper vs Our Setting
원 논문은 general scene의 multi-view (4+) generation을 대상으로 한다. 우리는 robotics dual-view (ego+wrist) 세팅에 맞게 핵심 컴포넌트를 교체한다.
Aspect Paper (VGGRPO) Our Version
Base model SV4D (multi-view) Wan2.1-T2V-1.3B (single→dual-view)
4D Recon DUSt3R / MASt3R Any4D (DINOv2 + DPT)
Connector Not specified Conv3d k=5 s=(1,2,2) 2.05M params
Domain General scenes Robotics (DROID ego+wrist)
GRPO base Custom implementation Flow-GRPO (existing, LoRA r=32)
Views 4+ multi-view 2-view (ego+wrist)

Technical Decisions with Rationale

Why Any4D instead of DUSt3R?
DUSt3R는 multi-view input을 요구하지만, 로봇 세팅에서는 ego+wrist 2 view만 존재한다. Any4D는 단일 비디오에서 4D 복원이 가능하며, DINOv2 backbone이 로봇 장면에서 검증되어 있다.
Why Flow-GRPO?
Wan2.1 네이티브 GRPO 지원으로 reward function 교체만으로 VGGRPO 통합이 가능하다. 별도 GRPO 구현 불필요.

Technical Challenges

Depth Scale Alignment
Euler Pose Format
2-view View Limitation
Gap Domain Shift

Details

Open Questions

References

VGGRPO — An et al., arXiv 2603.26599 (Google, U Copenhagen, Oxford)
Towards World-Consistent Video Generation with 4D Latent Reward. 본 연구의 source paper.
Any4D
Feed-forward 4D reconstruction (DINOv2 + DPT). DUSt3R 대체로 채택.
Flow-GRPO — github.com/yifan123/flow_grpo
Wan native GRPO implementation. GRPO optimization backbone으로 사용.
DUSt3R — Wang et al., CVPR 2024
Dense 3D reconstruction from stereo. 원 논문의 4D reconstruction backbone.
DROID — Khazatsky et al., RSS 2024
Robot manipulation dual-view dataset. 학습 데이터셋으로 활용.
Wan2.1 — arXiv 2503.20314
Video diffusion model (flow matching). Base video generator.