N2M: Bridging Navigation and Manipulation by Learning Pose Preference from Rollout

Determining where to execute the manipulation policy is a fundamental challenge in mobile manipulation. Most approaches have formulated this as a geometric search problem, prioritizing physical reachability. However, given the high sensitivity of modern learning-based manipulation policies, geometric criteria alone are insufficient. Optimal performance requires base positioning that is aware of the policy's preference. While recent works have attempted to address this, they remain limited in practicality due to reliance on pre-built scene reconstruction and slow inference. In this work, we introduce N2M that systematically reformulates the approach to base positioning problem, naturally overcoming limitations of previous methods. Our key insight is that policy preferences are inherent to the local scene structure and can be effectively learned from the policy rollouts. Technically, we propose a novel viewpoint augmentation strategy that enables the model to learn robust, viewpoint-invariant pose preferences with remarkable data efficiency. Extensive experiments demonstrate that N2M achieves state-of-the-art performance, outperforming both non-policy-aware baselines and recent policy-aware alternatives. Furthermore, we provide a comprehensive analysis highlighting N2M’s broad applicability, generalization capabilities, and data efficiency. Project website: https://clvrai.github.io/N2M/

When mobile robots perform tasks like opening a cabinet or picking up an item, they must first move into position before extending their robotic arms. Traditionally, engineers program robots to find any spot where the target object is physically within reach. However, modern AI-driven robots are incredibly sensitive; even if an object is technically reachable, the robot's onboard camera might view it from an unfamiliar angle, causing the task to fail completely. To solve this, we introduced N2M, a system that teaches a robot to predict the exact starting positions its AI arm "prefers" based on past practice. We developed a clever data-trick that allows the robot to simulate hundreds of different viewpoints on its own, vastly reducing the amount of real-world training data needed. Our experiments show that N2M allows robots to rapidly adapt to changes, such as a person shifting a chair in mid-motion, and smoothly complete complex, multi-step chores in real-world environments. Ultimately, this research provides a lightweight, plug-and-play enhancement that makes everyday smart robots much more reliable and practical for home and industrial use.