Research map / Detection and segmentation / Open-set and few-shot detection
Generalizing 6-DoF Grasp Detection via Domain Prior Knowledge
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.
Improves generalization of 6-DoF grasp detection to novel objects by adding physics-based antipodal regularization during training and contact-score joint test-time optimization with a projection contact map, evaluated on GraspNet-1Billion and a real UR-10 robot.
Problem
Learning-based 6-DoF grasp detectors perform well on seen and similar objects but degrade significantly on novel objects with diverse shapes and structures, and unstable grasps arise in cluttered scenes.
Approach
Injects domain prior knowledge of robotic grasping: a differentiable antipodal/collision/surface physical constraint regularizer computed from object SDF during training, plus test-time refinement of grasps by aligning a predicted contact map (with a novel projection contact distance) while constraining with a grasp score network.
Key results
- On the novel set of GraspNet-1Billion RealSense scenes, the full method reaches 36.67 AP versus 30.61 for the baseline and 28.06 for GSNet.
- Physical constraint regularization alone raises novel-set AP from 30.61 to 35.58 without hurting seen/similar sets (seen 66.67 vs 66.09).
- Real-world UR-10 experiments with 20 novel objects: isolated success rate 68.33% vs 58.33% baseline; cluttered success rate 64.86% vs 48.21%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add physical constraint regularization | Ours Baseline · GraspNet-1billion RealSense novel | AP: 30.61 -> 35.58 | Helped |
| Add contact-score joint optimization with the projection contact map and predicted mask | PCR model · GraspNet-1billion RealSense novel | AP: 35.58 -> 36.67 | Helped |
| Add physical constraint regularization | Ours Baseline · GraspNet-1billion RealSense seen | AP: 66.09 -> 66.67 | Helped |
| Add physical constraint regularization | Ours Baseline · GraspNet-1billion RealSense similar | AP: 64.82 -> 65.55 | Helped |
| Add antipodal regularization (ablation) | PCR · GraspNet-1billion RealSense novel | AP: 32.38 -> 35.58 | Helped |
| Add collision constraint (ablation) | PCR · GraspNet-1billion RealSense novel | AP: 32.34 -> 35.58 | Helped |
| Add surface constraint (ablation) | PCR · GraspNet-1billion RealSense novel | AP: 35.57 -> 35.58 | No clear effect |
| Use the original contact map instead of the projection contact map (ablation) | C-SJO · GraspNet-1billion RealSense novel | AP: 36.67 -> 36.27 | Hurt |
| Remove contact-map optimization (ablation) | C-SJO · GraspNet-1billion RealSense novel | AP: 36.67 -> 35.86 | Hurt |
| Remove score optimization (ablation) | C-SJO · GraspNet-1billion RealSense novel | AP: 36.67 -> 36.23 | Hurt |
| Add domain prior knowledge | Baseline · Real-world isolated objects | SR (%): 58.33 (35/60) -> 68.33 (41/60) | Helped |
| Add domain prior knowledge | Baseline · Real-world cluttered objects | SR (%): 48.21 (43/89) -> 64.86 (48/74) | Helped |
Surprising findings
- Object augmentation via merging similar-set objects improves novel AP only modestly (32.30) compared with the proposed prior-knowledge approach (36.67) and the two are complementary (37.00 combined).
- C-SJO has almost no effect on seen and similar sets, gains are concentrated on novel objects.
Limitations
- Gains on novel objects remain well below seen-set performance (36.67 vs 66.61 AP).
- Test-time optimization with Adam adds per-grasp iterative refinement cost and relies on noisy depth-derived point clouds.
- Relies on multi-view TSDF reconstruction of scenes, which may not be available in single-view settings.
Experimental setup
Datasets: GraspNet-1Billion
Benchmarks: GraspNet-1Billion
Metrics: AP AP0.8 AP0.4 success rate scene completion rate
Baselines: GraspNet-baseline Scale-balanced Grasp GSNet Object augmentation (merged similar set)
Builds on: GSNet Scale-balanced grasp detection GraspNet-1Billion KinectFusion Minkowski Engine
Closest papers by meaning
- Active Perception for Grasp Detection via Neural Graspness Field (NeurIPS 2024 · similarity 0.88)
- Grasp2Grasp: Vision-Based Dexterous Grasp Translation via Schrödinger Bridges (NeurIPS 2025 · similarity 0.86)
- Contact Map Transfer with Conditional Diffusion Model for Generalizable Dexterous Grasp Generation (NeurIPS 2025 · similarity 0.86)
- Grasp-Then-Plan with Failure Attribution: A Closed Two-Stage Framework for Precise and Generalizable Robotic Manipulation (NeurIPS 2026 · similarity 0.86)
- Vision Foundation Model Enables Generalizable Object Pose Estimation (NeurIPS 2024 · similarity 0.86)
- DexGrasp Anything: Towards Universal Robotic Dexterous Grasping with Physics Awareness (CVPR 2025 · similarity 0.85)
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.