Active Perception for Grasp Detection via Neural Graspness Field
Authors: Haoxiang Ma, Modi Shi, Boyang Gao, Di Huang
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on the Grasp Net-1Billion benchmark demonstrate significant performance improvements compared to previous works. Real-world experiments show that our method achieves a superior trade-off between grasping performance and time costs. |
| Researcher Affiliation | Collaboration | Haoxiang Ma1 Modi Shi1 Boyang Gao2 Di Huang1 1State Key Laboratory of Complex and Critical Software Environment, School of Computer Science and Engineering, Beihang University, Beijing, China 2Geometry Robotics |
| Pseudocode | No | The paper describes its methods using text and mathematical equations but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github. com/ mahaoxiang822/Active NGF. |
| Open Datasets | Yes | We construct a simulation active grasp benchmark based on the Grasp Net-1Billion benchmark [10], which consists of 100 scenes for training and 90 scenes for testing. |
| Dataset Splits | Yes | The test set is divided into seen, similar, and novel sets based on the included objects. |
| Hardware Specification | Yes | The training and evaluation of the simulation experiments are conducted on a single NVIDIA V100 GPU. We conduct real-world experiments of the proposed active grasp detection method on a 6-Do F UR-10 robot arm with a mounted Real Sense D435i depth camera. The analysis is performed on a workstation with a single NVIDIA 3090 GPU and an AMD Ryzen 5 2600 six-core processor. |
| Software Dependencies | No | The paper mentions specific frameworks like GSNet [31] and ESLAM [12] but does not provide specific version numbers for general software dependencies such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | For the mapping of NGF, the first view is trained for 100 iterations and following views are trained for 50 iterations. For each ray, 32 points are sampled for stratified sampling and 8 points for importance sampling. For NBV planning, we downsample the original image to 1/8 to sample rays for graspness rendering to speed up the computation of view information gain. For each scene, 1024 points are sampled from the NGF for grasp pose synthesis. |