DQ-Net · DQ-Bench

Whole-Body Coordination for Dynamic Object Grasping with Legged Manipulators (AAAI 2026)

Dynamic quadruped grasping with whole-body coordination, benchmarked on DQ-Bench and powered by DQ-Net.

Qiwei Liang, Boyang Cai, Rongyi He, Hui Li, Tao Teng, Haihan Duan, Changxin Huang, Runhao Zeng*

Hong Kong University of Science and Technology (Guangzhou), Shenzhen University,
T-Stone Robotics Institute (CUHK), Shenzhen MSU-BIT University

📄 Paper (PDF) 💻 arXiv 🧩 Code
Robotics · Legged Manipulation Dynamic Grasping Teacher–Student Learning

Abstract

Overview

We introduce DQ-Bench, the first benchmark for dynamic object grasping with quadruped robots, supporting realistic dynamics, diverse objects, multi-level task difficulty, and comprehensive evaluation. Building upon this benchmark, we propose DQ-Net, a teacher–student framework combining a Grasp Fusion Module and lightweight dual-view student network for stable and efficient whole-body dynamic grasping. Extensive experiments show DQ-Net outperforms baselines in both success rate and responsiveness.

DQ-Bench

Benchmark
  • Built on Isaac Gym for high-performance simulation.
  • Four difficulty levels: Low-Speed 2D → High-Speed 3D with rough terrain.
  • Diverse seen/unseen YCB objects for generalization testing.
  • Evaluation metrics: GSR, OSSR, TSC.
DQ-Bench Task Levels

DQ-Net Framework

Method

DQ-Net integrates a Grasp Fusion Module (GFM) with a hierarchical teacher–student structure:

  1. Teacher: Uses privileged info (pose, velocity, point cloud) + GFM to output optimal grasp poses.
  2. Student: Lightweight Transformer-based network with dual-perspective depth & mask inputs.
  3. Low-level controller: Ensures coordinated locomotion and manipulation.
DQ-Net Architecture

Results

Experiments

DQ-Net achieves the highest grasp success rates across all difficulty levels and unseen object categories.

Grasp Success Rate Table
Grasp Success Rate (GSR) across difficulty levels on DQ-Bench.
Unseen Objects Performance
Generalization results on unseen YCB objects.
Qualitative Comparison
Qualitative comparison between DQ-Net and baseline methods.

Video

Demo

Citation

BibTeX 
@misc{liang2025wholebodycoordinationdynamicobject,
  title         = {Whole-Body Coordination for Dynamic Object Grasping with Legged Manipulators},
  author        = {Qiwei Liang and Boyang Cai and Rongyi He and Hui Li and
                   Tao Teng and Haihan Duan and Changxin Huang and Runhao Zeng},
  year          = {2025},
  eprint        = {2508.08328},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2508.08328}
}