Research map / 3D vision / 3D generation
Active Perception for Grasp Detection via Neural Graspness Field
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an active grasp detection framework that builds an online Neural Graspness Field (NGF) from camera views, plans next-best-views via graspness inconsistency, and samples grasps from the NGF.
Problem
Grasp detection in cluttered scenes suffers from incomplete 3D geometry from single views, while multi-view scanning is time-consuming; active view planning must balance grasp performance and camera movement time.
Approach
Model the scene's grasp distribution as a neural field (graspness) optimized online with rendering loss from a pre-trained graspness network, and select next-best views by the inconsistency between rendered graspness and pseudo-graspness from rendered depth, plus sample grasp positions decoded from the field.
Key results
- After 10 views, the method reaches 55.12 AP on GraspNet seen, 52.85 on similar and 24.74 on novel (RealSense), improving over ACE-NBV by about 8.4%, 6.7% and 3.0% respectively.
- Real-world grasp success rate of 74.67% (56/75), vs 70.67% for Close-loop NBV and 62.67% for ACE-NBV.
- Average runtime per planning step is 3.44 s, with robot execution about 51% of the time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use graspness inconsistency-guided next-best-view planning rather than ACE-NBV planning | ActiveNGF · GraspNet-1Billion seen (RealSense) | AP: +8.38% | Helped |
| Use graspness inconsistency-guided next-best-view planning rather than ACE-NBV planning | ActiveNGF · GraspNet-1Billion similar (RealSense) | AP: +6.71% | Helped |
| Use graspness inconsistency-guided next-best-view planning rather than ACE-NBV planning | ActiveNGF · GraspNet-1Billion novel (RealSense) | AP: +2.98% | Helped |
| Use ActiveNGF instead of ACE-NBV | ActiveNGF · Real-world grasping | Success Rate (%): 4.00% | Helped |
| Use ActiveNGF instead of Close-loop NBV | ActiveNGF · Real-world grasping | Success Rate (%): 12.00% | Helped |
| Add neural graspness sampling from NGF instead of sampling graspness from 3D geometry (ablation) | Grasp detection using ActiveNGF · GraspNet-1Billion seen, similar, and novel sets | AP | Helped |
| Select views using graspness inconsistency rather than geometry or appearance uncertainty | NGF-based next-best-view planning · GraspNet-1Billion seen, similar, and novel sets | AP | Helped |
| Replace inconsistency-guided planning with ACE-NBV affordance-based planning | NGF-based next-best-view planning · GraspNet-1Billion as views are added | AP | Hurt |
Surprising findings
- ACE-NBV achieves results comparable to the proposed method in initial steps but shows little improvement as more views are added.
Limitations
- NBV planning time grows with the number of candidate views.
- Cannot handle dynamic scene changes; failed grasps require re-running perception.
- Simulation benchmark is built on a public dataset, and code is promised only after acceptance.
Experimental setup
Datasets: GraspNet-1Billion
Benchmarks: GraspNet-1Billion (seen, similar, novel) real-world grasping on UR-10 with 25 YCB objects
Metrics: AP AP0.8 AP0.4 grasp success rate
Baselines: Close-loop NBV ACE-NBV ActiveNeRF Uncertainty-policy
Builds on: GSNet ESLAM StyleSDF Close-loop NBV ACE-NBV
Closest papers by meaning
- Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge (CVPR 2024 · similarity 0.88)
- Language-driven Grasp Detection (CVPR 2024 · similarity 0.85)
- CADGrasp: Learning Contact and Collision Aware General Dexterous Grasping in Cluttered Scenes (NeurIPS 2025 · similarity 0.83)
- ZeroGrasp: Zero-Shot Shape Reconstruction Enabled Robotic Grasping (CVPR 2025 · similarity 0.83)
- Grasp2Grasp: Vision-Based Dexterous Grasp Translation via Schrödinger Bridges (NeurIPS 2025 · similarity 0.83)
- DexFuncGrasp: A Robotic Dexterous Functional Grasp Dataset Constructed from a Cost-Effective Real-Simulation Annotation System (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.