LMGN: A Unified Policy Learning Model for Last-Mile Grasp-Navigation

Xuenmin Hu, Qingxue Qian, Wei Luo, Benyi Ma, Bo Tang, Long Chen

November 2026

TL;DR LMGN trains one policy for a quadruped with an arm that turns candidate grasps into where-to-stand and which-way-to-approach cues and switches between stage-specific experts. In Habitat it reaches 74.8% grasp success, against 64.5% for GAMMA and 52.2% for VBC.

What they do

A quadruped with an arm that reaches an object still has to decide where to stand, which side to approach from and when to move the arm instead of the base. LMGN learns this “last-mile” step as one policy, trained with PPO on top of the Visual Whole-Body Control (VBC) setup, for a Unitree B1 with a Z1 arm (19 DoF).

How it works

ED-DR features. GSNet proposes 6D grasps from the object point cloud. Each is converted into grasp height, approach direction and the base position that grasp implies, so the policy gets an explicit cue for where to stand.
Hierarchical mixture-of-experts. High-level experts cover approach and base adjustment, pre-grasp alignment and grasp, and lift or retry; low-level experts handle variation within a stage. Top-1 routing picks a single route per step.
Joint optimization. Navigation and grasp-reachability rewards, a contrastive loss that ties expert routes to stages, and a stage-weighted loss that keeps base, arm and gripper commands consistent.

Results

Habitat (GAMMA protocol): 73.9% gaze and 74.8% grasp success, against 69.0% and 52.2% for VBC and 66.3% and 64.5% for GAMMA.
Isaac Gym, six objects: 84.47% gaze and 53.86% grasp success, against 44.34% and 19.84% for VBC. Without joint optimization these drop to 26.67% and 2.79%; without the consistency loss alone, to 19.10% and 6.59%.
Three unseen objects: 77.49% gaze and 48.52% grasp success with no fine-tuning.

Limitations

All experiments are in simulation; no real-robot results are reported. Grasp success in Isaac Gym stays near 54%, and the gain over the variant without the expert hierarchy is small there (53.86% vs 50.07%). On the bowl, gaze success is 38.08%, below that variant’s 49.53%.

Figures

From Figure 1: LMGN fuses grasp-guided reachability features, the environment point cloud, the object pose and propri... — From Figure 1: LMGN fuses grasp-guided reachability features, the environment point cloud, the object pose and proprioception; a hierarchical mixture-of-experts picks a stage-specific route that outputs base velocity, end-effector pose change and gripper command.

From Figure 2: The ED-DR representation turns a 6D grasp pose into grasp height, approach direction and the relative ... — From Figure 2: The ED-DR representation turns a 6D grasp pose into grasp height, approach direction and the relative base position from which to grasp.

From Figure 3: Weights of the three high-level experts over one episode, with snapshots of approach, pre-grasp alignm... — From Figure 3: Weights of the three high-level experts over one episode, with snapshots of approach, pre-grasp alignment, grasp and lift.

Abstract

Legged robots equipped with manipulators have shown significant potential in the fields of daily life and industry in complex environments. However, existing methods mainly focus on global navigation or manipulation independently, while overlooking the last-mile navigation problem, where the robot must jointly determine the base placement, approach direction, and grasp execution to achieve reliable object grasping, and different manipulation stages require distinct control behaviors. To address these challenges, we propose Last-Mile Grasp-Navigation (LMGN), a unified policy learning model for Quadruped robot loco-manipulation. An Explicit Distance and Directional Reachability (ED-DR) representation is first introduced to explicitly represent the features of grasp-compatible base placement and approach direction. We then develop a semantically aligned Hierarchical Mixture-of-Experts (HMoE) policy that models stage-aware behaviors across the stages of base adjustment, pre-grasp alignment and grasp execution, and object lift. Finally, a new optimization method that jointly aligns navigation and manipulation, stage-dependent expert routing, and action-level consistency is proposed to optimize the HMoE policy. Extensive experiments on the Habitat and Isaac Gym simulators demonstrate that LMGN consistently outperforms state-of-the-art methods, denoting that integrating grasp-aware reachability with stage-aware policy learning significantly improves the success rates of last-mile navigation and target grasp for robots.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong