Reproduce “DVDP: An End-to-End Policy for Mobile Robot Visual Docking with RGB-D Perception” 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
Haohan Min, Zhoujian Li, Yu Yang, Jinyu Chen, Shenghai Yuan
Researchers from Tsinghua University, National University of Singapore, and Nanyang Technological University developed DVDP, an end-to-end deep learning framework for robust mobile robot visual docking using RGB-D perception. This policy directly generates precise docking trajectories from diverse initial poses, achieving a 73.2% success rate and outperforming other approaches by a substantial margin.
Key insights
- The 'hierarchical-then-contextual' feature fusion (DLA backbone + Pyramid Pooling Module) and an MLP-based non-linear cross-attention mechanism are crucial for effectively integrating multimodal RGB and depth features.
- An autoregressive dual-decoder for trajectory generation is essential for modeling temporal dependencies, leading to smoother, kinematically feasible robot movements.
- Training on a diverse hybrid dataset, combining virtual and real-world environments with domain randomization, is effective for generalizing end-to-end policies for complex robotic tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.