ambitious-octopus / agile_grasp2

A ROS package to detect grasp poses in point clouds.

agile_grasp 2.0

• Author: Andreas ten Pas ([email protected])
• Version: 1.0.0
• ROS Wiki page: not available yet
• Author's website: http://www.ccs.neu.edu/home/atp/

1) Overview

This package localizes antipodal grasps in 3D point clouds. AGILE stands for Antipodal Grasp Identification and LEarning. The reference for this package is: High precision grasp pose detection in dense clutter. agile_grasp 2.0 is an improved version of our previous package, agile_grasp.

The package already comes with a pre-trained machine learning classifier and can be used (almost) out-of-the-box with RGBD cameras such as the Microsoft Kinect and the Asus Xtion Pro as well as point clouds stored as *.pcd files.

If you like this package and use it in your own work, please cite our arXiv paper.

2) Requirements

1. ROS Indigo (installation instructions for Ubuntu)
2. OpenNI or a similar range sensor driver
3. Caffe

3) Installation

1. Install Caffe. Instructions for Ubuntu 14.04.

2. Compile Caffe as a cmake library (https://github.com/BVLC/caffe/pull/1667):

   $ cd caffe && mkdir cmake_build && cd cmake_build
   $ cmake .. -DBUILD_SHARED_LIB=ON
   $ make -j 12
   $ ln -s ../build .
   

   If the second line gives the error Manually-specified variables were not used by the project: BUILD_SHARED_LIB, just run the second line again.

3. Clone the agile_grasp 2.0 repository into your ros workspace:

   $ cd <location_of_your_ros_workspace>/src
   $ git clone https://github.com/atenpas/agile_grasp2
   

4. Recompile your ROS workspace:

   $ cd ..
   $ catkin_make
   

4) Detect Grasp Poses With a Robot

1. Connect a range sensor, such as a Microsoft Kinect or Asus Xtion Pro, to your robot.

2. Adjust the file robot_detect_grasps.launch for your sensor and robot hand.

3. Run the grasp pose detection:

   $ roslaunch agile_grasp2 robot_detect_grasps.launch
   

5) Detect Grasp Poses in a PCD File

1. Have a *.pcd file available. Let's say it's located at /home/user/cloud.pcd.

2. Change the parameter cloud_file_name in the file file_detect_grasps.launch to /home/user/cloud.pcd.

3. Detect grasp poses:

   roslaunch agile_grasp2 file_detect_grasps.launch
   

6) Using Precomputed Normals

The ROS node, grasp_detection, can take point clouds with normals as input. The normals need to be stored in the normal_x, normal_y, and normal_z fields. Attention: the only preprocessing operation that works with this is workspace filtering, so avoid voxelization and manual indexing.

7) Parameters

The most important parameters are cloud_type and cloud_topic which define the type of the input point cloud and from which topic the cloud is received, and workspace and num_samples which define the workspace dimensions and the number of grasp hypotheses to be sampled. The other parameters only need to be modified in special cases.

Input

• cloud_type: the type of the input point cloud.
  • 0: *.pcd file
  • 1: sensor_msgs/PointCloud2
  • 2: agile_grasp/CloudSized
  • 3: agile_grasp/CloudIndexed.
• cloud_topic: the ROS topic from which the input point cloud is received.
• samples_topic: the ROS topic from which the input samples are received.

Visualization

• plot_mode: what type of visualization is used (0: no visualization, 1: PCL, 2: rviz (not supported yet))
• only_plot_output: set this to false to see additional visualizations.

Grasp Hypothesis Search

• workspace: the limits of the robot's workspace, given as a cube centered at the origin: [x_min, x_max, y_min, y_max, z_min, z_max] (in meters).
• camera_pose: the camera pose as a 4x4 homogeneous transformation matrix.
• num_samples: the number of grasp hypotheses to be sampled.
• num_threads: the number of CPU threads to be used.
• nn_radius_taubin: the neighborhood search radius for normal estimation (in meters).
• nn_radius_hands: the neighborhood search radius for the grasp hypothesis search (in meters).
• num_orientations: the number of hand orientations to be considered.
• antipodal_mode: the output of the algorithm. 0: grasp hypotheses, 1: antipodal graps (prediction), 2: antipodal grasps (geometric).
• voxelize: if the point cloud gets voxelized.
• filter_half_grasps: if half-grasps are filtered out.
• gripper_width_range: the aperture range of the robot hand: [aperture_min, aperture_max].

Robot Hand Geometry

• finger_width: the finger width.
• hand_outer_diameter: the hand's outer diameter (aperture + finger_width).
• hand_depth: the hand's depth (length of fingers).
• hand_height: the hand's height.
• init_bite: the initial amount that the hand extends into the object to be grasped.

Classifier

• model_file: the Caffe prototxt file that specifies the network.
• trained_file: the Caffe model file that contains the weights for the network.
• label_file: a txt file that contains the label for each class.
• min_score_diff: the minimum difference between the positive and the negative score for a grasp to be classfied as positive.
• batch_size: the number of images used per batch in Caffe (larger batch size => faster, but larger memory usage)

Clustering

• min_inliers: the minimum number of grasps required to form a cluster.

Grasp Selection

• num_selected: the number of selected grasps. If antipodal grasps are predicted/calculated, then the selected grasps will be the highest-scoring grasps.

Other

• use_service: uses a ROS service instead of topics (untested).

8) Citation

If you like this package and use it in your own work, please cite our arXiv paper:

@misc{1603.01564,
Author = {Marcus Gualtieri and Andreas ten Pas and Kate Saenko and Robert Platt},
Title = {High precision grasp pose detection in dense clutter},
Year = {2016},
Eprint = {arXiv:1603.01564},
}