robotology/skeleton3D

Bridge between 2D Tensorflow-based human pose estimation and 3D estimation from stereovision

skeleton3D

Introduction

Bridge between 2D human pose estimation and 3D estimation from stereovision

Table of Contents

1. Dependencies
2. Results
3. Build
4. How-to
5. References

Dependencies

• YARP
• iCub
• icub-contrib-common
• icub-hri
• objectsPropertiesCollector (OPC): a robot's working memory.
• stereo-vision: 3D estimation from stereo-vision camera
• deeperCut based skeleton2D or yarpOpenPose: 2D human pose tracking
• Optional modules: for human-robot interaction demos
  • Peripersonal Space
  • react-ctrl
  • modified onthefly-recognition: the built module's name is changed, so there will be no conflict with the official module.
  • cardinal-points-grasp

Results

Safe HRI demo

Hand-over demo

Build

Build and Install normally, i.e.

mkdir build && cd build
ccmake ..
make install

How-to

Safe HRI demo

1. Open the application with openpose, PS_modulation_iCub_skeleton3D_openpose, or application with deepcut, PPS_modulation_iCub_skeleton3D in yarpmanager. Note that application with deeperCut provides more responsive robot's actions.
2. Launch all module and connect.
3. (Optional) If you want to use application with deeperCut, you have to run skeleton2D.py in terminal rather than yarpmanager. The possibility to run python script from yarp manager is broken now.

   # Open a terminal and ssh to machine with GPU, e.g. `icub-cuda`
   ssh icub-cuda
   skeleton2D.py --des /skeleton2D --gpu 0.7
   

4. Users can log into rpc service of the module to set the parameters by:

   yarp rpc /skeleton3D/rpc
   # help function by typing:
   help
   

5. Move the icub's neck to look down about 23 degree, e.g. with yarpmotorgui. If you run icubCollaboration (see below), this step is not necessary.
6. Connect to the rpc service of react-controller, and make the controlled arm (left by default) move:
   • To a fix position: in this mode, robot tries to keep its end-effector at a fix position, e.g. (-0.3,-0.15,0.1) for left_arm of icub, while avoiding human's body parts

   yarp rpc /reactController/rpc:i  
   # for the *left_arm*
   set_xd (-0.3 -0.15 0.1)
   
   # or for the *right_arm*
   set_xd (-0.3 0.15 0.1)
   
   # to stop typing:
   stop
   

   • In a circle: in this mode, robot moves its end-effector along a circle trajectory in the y and z axes, relative to the current end-effector position, while avoiding human's body parts. The first command moves robot's arm to a tested safe initial position for the circle trajectory.

   set_xd (-0.3 -0.15 0.1)
   set_relative_circular_xd 0.08 0.27
   
   # to stop typing:
   stop
   

• Note: users can tune the workspace parameters in configuration file to constrain the robot's partner. The module currently works with only one partner at a time.

Hand-over demo

1. First, do all the above step
2. Open the application script, ontheflyRecognition_PPS_both, in yarpmanager. This app allows on-hand object training and on-hand object recognition.

   # Connect to **skeleton3D**:
   yarp rpc /skeleton3D/rpc
   enable_tool_training right
   
   # Connect to **onTheFlyRecognition_right**
   yarp rpc /onTheFlyRecognition_right/human:io
   # Hold object on the right hand and type:	
   train <object_name> 0
   
   # The whole procedure can be applied for the left hand also
   

3. Open the application script, iolVM_Phuong, in yarpmanager. This app allows on-table object recognition for grasping
4. Open the application script, grasp-processor, in yarpmanager. This app allows robot to grasp recognized object on the table.
5. Run module icubCollaboration. Currently, all connections to other modules are internally, so it needs to run after all others.
6. Connect all ports.

   # the robot arm using for **icubCollaboration** needs to be the same as **react-ctrl** above
   icubCollaboration --robot icub --part <right_arm/left_arm>
   
   # rpc access to the module
   yarp rpc /icubCollaboration/rpc
   
   # type help for all support commands
   help
   
   # hold a trained object (within the robot's reachable area) and type:
   receive <object_name> 
   # robot should detect the object, take-over it and put it on the table (see the video)
   
   # ask robot to give the object on the table
   pre_grasp_pos
   hand_over_object <object_name> <handRight/handLeft>
   
   

References

D. H. P. Nguyen, M. Hoffmann, A. Roncone, U. Pattacini, and G. Metta, “Compact Real-time Avoidance on a Humanoid Robot for Human-robot Interaction,” in Proceedings of the 2018 ACM/IEEE International Conference on Human-Robot Interaction, 2018, pp. 416–424.

P. D. Nguyen, F. Bottarel, U. Pattacini, H. Matej, L. Natale, and G. Metta, “Merging physical and social interaction for effective human-robot collaboration,” in Humanoid Robots (Humanoids), 2018 IEEE-RAS 18th International Conference on, 2018, pp. 710–717.