Reproduce “N2M: Bridging Navigation and Manipulation by Learning Pose Preference from Rollout” 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
Kaixin Chai, Hyunjun Lee, Joseph J. Lim
Researchers at KAIST and Vital Robotics developed N2M, a system that learns optimal base pose preferences for mobile manipulators directly from policy rollouts, bridging navigation and manipulation. This approach achieves higher success rates (e.g., 0.55-0.56 vs. 0.09-0.24 for a baseline) with rapid inference (0.07 seconds) and strong generalization across various tasks and unseen environments using limited data.
Key insights
- Learning policy preferences directly from a limited number of successful rollouts, rather than relying on geometric reachability or heuristic policy-awareness, enables mobile manipulators to consistently find optimal base poses for diverse manipulation tasks.
- A key data augmentation technique, viewpoint augmentation, significantly improves data efficiency and allows the model to generalize robustly to unseen environments and varying initial robot viewpoints using minimal real-world data.
- Predicting a multi-modal distribution of preferable poses (e.g., using a GMM) from a single ego-centric observation enables the system to capture multiple valid task-execution viewpoints and achieve rapid, real-time inference (0.07 seconds per inference).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.