Hang Yin

Robotics @ Stanford

This project is meant to turn a Franka Emika Panda arm into a Jenga assistant! It uses computer vision to detect Jenga bricks and place them on top of the tower. The robot plans and executes trajectories using a custom MoveGroup API.

Video Demo

Calibration

  • Calibrating the robot relies on adding frames to the tf tree based on known geometry.
  • The robot is sent to a calibration position, then given an april tag that aligns with the gripper frame.
  • Once the camera can see the tag, it creates series of transforms from the camera to the tag and finally to the base of the robot.
  • Calibration pose:

Calibration Pose

Tower Detection

  • We detect the tower using the depth image from an Intel Realsense camera.
  • Once the top of the tower and the table are found by generating contours on the depth image, we repeatedly scan between these locations looking for partially removed pieces.
  • The starting orientation of the tower is determined through line detection in OpenCV.

Tower Detection

Machine Learning for Hand Detection

  • Overview:
    • We performed transfer learning on a deep neural network to help us detect whether there are hands in the scene so that the robot knows when it should look for a Jenga brick.
  • Dataset:
    • Video samples of the Jenga tower were taken and separated into two classes: with hand(s) and without hand(s) around.
    • 1000 images were extracted from each class.
  • Model:
    • We performed transfer learning on a pretrained neural network - MobileNets
    • The MobileNet we are using implements depth-wise separable convolutions to build light weight deep neural networks
    • We only train the last layer of the MobileNet to perform our specific task
  • Training:
    • Epochs: 50
    • Batch Size: 16
    • Learning Rate: 0.001

Custom MoveIt Library

  • Overview:
    • This package allows us to move our robot around while also avoiding objects by using different services.
  • GetCartesianPath:
    • Moves the end-effector gripper in a straight line from start to end.
    • Uses multiple intermediate waypoints to ensure straight path.
  • GetPositionIK:
    • Calculates possible joint positions of robot from a given start and end position and orientation for the end-effector.
    • Used to Orient the gripper without shifting the position .
    • Used to move the robot when straight movement is not necessary.
  • GetPlanningScene:
    • The environment around the robot.
    • This allows us to add in the table, camera, and jenga tower to avoid collisions .

MoveIt Visual

Source code

Github repo

Reference

  • Howard, A., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., & Adam, H. (2017). MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. https://doi.org/10.48550/ARXIV.1704.04861

Group members

Katie Hughes, Liz Metzger, Alyssa Chen

Updated: