Reproduce “OFlow: Injecting Object-Aware Temporal Flow Matching for Robust Robotic Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kuanning Wang, Ke Fan, Chenhao Qiu, Zeyu Shangguan, Yuqian Fu, Yanwei Fu, Daniel Seita, Xiangyang Xue
OFlow injects object-aware temporal flow matching into Vision-Language-Action (VLA) models, unifying foresight and object reasoning within a shared semantic latent space for robotic manipulation. The framework achieves 96.6% success on the LIBERO benchmark and an average of 69% on real-world tasks, demonstrating enhanced robustness to dynamic environments and distribution shifts.
Key insights
- Unifying object-aware perception and temporal foresight within a single semantic latent space (DINOv2 features) significantly enhances VLA model robustness and computational efficiency.
- Predicting abstract semantic states rather than pixel-level future frames provides a more stable and effective form of foresight, reducing brittleness to visual noise and improving generalization.
- Self-supervised object decomposition using learning-free clustering on latent features allows for autonomous extraction of task-relevant information without explicit annotations, making the system more adaptable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.