In this topic, the purpose is to realize the system of robotic platform based on deep learning. As the main research method of this topic, the Generative Adversarial Network  is used to establish the kinematics of robots, and the advantages and disadvantages between GAN and traditional mainstream models analyzed by kinematics.
In terms of process, we’ve divided the process into four parts, namely robotics platform hardware, control, data collection, and deep learning to building kinematics.  The self-built YOLO v3 model is combined with built-in IR projector of Kinect v2 to detect the object, so that the robot can detect an object in our platform. Finally, the robot uses the space of object to coordinate as the input of our GAN motion model, and outputs the joint angle of robot arms. The robotic arm can move to the object and grab it to complete the mission.

In this topic, we use the deep learning architecture to train and build a set of reverse kinematics models. Combining with an object recognition model at the same time, so that the robot has a certain item recognition ability to solve the error output caused by traditional inverse kinematics and data unevenness.

When the robotic arm performs a task, it often encounters the problem of overcoming inverse kinematics. In the real world, the traditional inverse kinematics control arm will affect the overall output accuracy due to the poor environment or current state.
This topic will study aiming at the robot arm with reasonable motion data, directly collecting the motion data on each arm structure, training and building a set of reverse kinematics models and adding it to another object recognition model, so that the robot has object recognition ability.