How to install GPD on Ubuntu 20.04 or newer?

Follow the Ubuntu 16.04 instructions, but you may need to adjust dependency versions and compiler flags, as the README doesn't provide updated guidance for newer distributions. Expect potential conflicts with PCL or OpenCV packages.

GPD vs. Dex-Net: which is better for cluttered environments?

GPD is specialized for 6-DOF grasp detection in dense clutter using point clouds, while Dex-Net often focuses on 2.5D depth images and may offer faster inference. Choose GPD for full orientation needs and proven clutter performance from its research papers.

Can I use GPD with a custom depth sensor not listed in the README?

Yes, as long as you can provide point clouds in PCD or PLY format. You'll need to ensure the data is registered and the camera position is correctly specified in the configuration file for normal estimation.

How to train a custom neural network for GPD using my own dataset?

The README outlines steps with PyTorch: compile with BUILD_DATA_GENERATION flag, generate grasp images, resize HDF5 datasets, train the network, and convert to ONNX then OpenVINO format. It requires OpenCV Contribs and careful parameter tuning.

Is GPD suitable for real-time industrial automation?

It can be optimized with OpenVINO for speed, but the sampling and classification steps may still be computationally heavy. Test thoroughly on your hardware, as real-time performance depends on point cloud density and parameter settings.

What are the hardware requirements for running GPD with GPU acceleration?

For GPU use with Caffe or OpenVINO, you'll need a compatible NVIDIA GPU and proper drivers. The Eigen-based CPU backend is lighter but slower; refer to the CNN Frameworks section for backend-specific details.

Open-Awesome

gpd

BSD-2-ClauseC++2.0.0

Detects 6-DOF grasp poses for parallel jaw grippers in 3D point clouds, enabling robotic grasping of novel objects in clutter.

GitHub

741 stars246 forks0 contributors

What is gpd?

Grasp Pose Detection (GPD) is a software package that detects 6-degree-of-freedom grasp poses for parallel jaw grippers in 3D point clouds. It solves the problem of robotic grasp planning for novel objects in cluttered environments by sampling grasp candidates and classifying them using neural networks. The system outputs viable grasp positions and orientations without requiring prior object models.

Target Audience

Robotics researchers and engineers working on robotic manipulation, grasp planning, and perception systems, particularly those using depth sensors and point cloud data.

Value Proposition

Developers choose GPD for its proven performance in clutter, support for multiple neural network backends (OpenVINO, Caffe, custom CPU implementation), and ability to handle novel objects without CAD models. It provides a complete, research-validated solution for 6-DOF grasp detection.

Overview

Detect 6-DOF grasp poses in point clouds

Use Cases

Best For

Robotic grasp planning for unknown objects in cluttered environments
Research projects requiring 6-DOF grasp pose detection from point clouds
Integration with ROS-based robotic manipulation systems
Benchmarking grasp detection algorithms against published results
Educational purposes in robotics perception and manipulation courses
Deploying grasp detection on various hardware using OpenVINO or Caffe backends

Not Ideal For

Projects requiring grasp detection for multi-fingered or dexterous hands beyond parallel jaw grippers.
Teams needing a containerized, plug-and-play solution without compiling C++ dependencies from source.
Applications where real-time performance on low-power embedded systems is critical, as GPD's neural network backends can be resource-intensive.

Pros & Cons

Pros

Clutter-Robust Detection

Designed specifically for dense cluttered environments, GPD effectively samples and classifies grasps even when objects are closely packed, as highlighted in its key features.

Flexible Neural Network Backends

Supports multiple frameworks including OpenVINO for speed, Caffe for GPU acceleration, and a custom Eigen-based CPU implementation, allowing deployment on various hardware (CNN Frameworks section).

Novel Object Compatibility

Works without pre-existing CAD models by directly processing point cloud data, enabling grasping of unseen objects, a core strength stated in the README.

Multi-View Sensor Support

Accommodates single or dual depth sensor configurations with pre-trained models for different view angles, improving scene perception as described in the Views section.

Cons

Complex Dependency Management

Requires specific versions of PCL, Eigen, OpenCV, and optional frameworks like OpenVINO or Caffe, making installation non-trivial and error-prone, especially on non-Ubuntu systems.

Limited to Parallel Jaw Grippers

Focused exclusively on two-finger robot hands, so it cannot handle more complex gripper types without significant modification, limiting its applicability.

Documentation Tied to Older Systems

Installation instructions are primarily tested on Ubuntu 16.04, and troubleshooting tips suggest frequent build issues, indicating portability challenges and outdated guidance.

Frequently Asked Questions

Related Projects

GitHub repository

Open3D: A Modern Library for 3D Data Processing

Stars13,552

Forks2,560

Last commit7 days ago

Draco

Draco is a library for compressing and decompressing 3D geometric meshes and point clouds. It is intended to improve the storage and transmission of 3D graphics.

Stars7,253

Forks1,051

Last commit27 days ago

mmdetection3d

OpenMMLab's next-generation platform for general 3D object detection.

Stars6,403

Forks1,758

Last commit1 year ago

PCDet

OpenPCDet Toolbox for LiDAR-based 3D Object Detection.

Stars5,557

Forks1,447

Last commit6 months ago

Community-curated · Updated weekly · 100% open source

Found a gem we're missing?

Open-Awesome is built by the community, for the community. Submit a project, suggest an awesome list, or help improve the catalog on GitHub.

Submit a project Star on GitHub

gpd