stretch_deep_perception provides demonstration code that uses open deep learning models to perceive the world.
This code depends on the stretch_deep_perception_models repository, which should be installed under ~/stretch_user/ on your Stretch RE1 robot.
Link to the stretch_deep_perception_models repository: https://github.com/hello-robot/stretch_deep_perception_models
There are four demonstrations for you to try.
First, try running the face detection demonstration via the following command:
roslaunch stretch_deep_perception stretch_detect_faces.launch
RViz should show you the robot, the point cloud from the camera, and information about detected faces. If it detects a face, it should show a 3D planar model of the face and 3D facial landmarks. These deep learning models come from OpenCV and the Open Model Zoo (https://github.com/opencv/open_model_zoo).
You can use the keyboard_teleop commands within the terminal that you ran roslaunch in order to move the robot's head around to see your face.
i (tilt up)
j (pan left) l (pan right)
, (tilt down)
Pan left and pan right are in terms of the robot's left and the robot's right.
Now shut down everything that was launched by pressing q and Ctrl-C in the terminal.
Second, try running the object detection demo, which uses the tiny YOLO v3 object detection network (https://pjreddie.com/darknet/yolo/). RViz will display planar detection regions. Detection class labels will be printed to the terminal.
roslaunch stretch_deep_perception stretch_detect_objects.launch
Once you're ready for the next demo, shut down everything that was launched by pressing q and Ctrl-C in the terminal.
Third, try running the body landmark point detection demo. The deep learning model comes from the Open Model Zoo (https://github.com/opencv/open_model_zoo). RViz will display colored 3D points on body landmarks. The network also provides information to connect these landmarks, but this demo code does not currently use it.
roslaunch stretch_deep_perception stretch_detect_body_landmarks.launch
Once you're ready for the next demo, shut down everything that was launched by pressing q and Ctrl-C in the terminal.
Finally, try running the nearest mouth detection demo. RViz will display a 3D frame of reference estimated for the nearest mouth detected by the robot. Sometimes the point cloud will make it difficult to see. Disabling the point cloud view in RViz will make it more visible.
We have used this frame of reference to deliver food near a person's mouth. This has the potential to be useful for assistive feeding. However, use of this detector in this way could be risky. Please be very careful and aware that you are using it at your own risk.
A less risky use of this detection is for object delivery. stretch_demos has a demonstration that delivers an object based on this frame of reference by holding out the object some distance from the mouth location and below the mouth location with respect to the world frame. This works well and is inspired by similar methods used with the robot EL-E at Georgia Tech [1].
roslaunch stretch_deep_perception stretch_detect_nearest_mouth.launch
Underlying the deep perception demos provided in this package is the OpenCV deep learning submodule, dnn, which is backed by the Intel Inference Engine. A python package called opencv-python-inference-engine compiles the inference engine into OpenCV, provided as cv2.dnn. Unfortunately, this package conflicts with the opencv-contrib-python package, which provide additional submodules, such as ARUCO detection support. This is why a version of the opencv-python-inference-engine was recompiled under a different name, renamed-opencv-python-inference-engine. This package is set up by default on the Stretch robot as of June 2022. With this new package, the deep learning submodule is provided as renamed_cv2.dnn. For developers looking to the source code in this repo as a guide, please take note of the OpenCV import statement which aliases the renamed OpenCV module to the more commonly seen module name: import renamed_cv2 as cv2. When you'd like to use the contrib modules from OpenCV (e.g. cv2.aruco), you can import OpenCV as per usual: import cv2. Please feel free to ask any questions on the forum.
[1] Hand It Over or Set It Down: A User Study of Object Delivery with an Assistive Mobile Manipulator, Young Sang Choi, Tiffany L. Chen, Advait Jain, Cressel Anderson, Jonathan D. Glass, and Charles C. Kemp, IEEE International Symposium on Robot and Human Interactive Communication (RO-MAN), 2009. http://pwp.gatech.edu/hrl/wp-content/uploads/sites/231/2016/05/roman2009_delivery.pdf
For license information, please see the LICENSE files.