Robot arm simulation based on model-free reinforcement learning

Regarding the control grasping operation of the robotic arm as the research content, for objects of arbitrary shape, the observation image is used as the input data, and the continuous action space is the output. Deep reinforcement learning is used to carry out research on the grasping control process of the object, so as to realize the transformation from high-dimensional images to end-to-end output of control commands. The article will design to build a simulated robotic arm environment.to use the bullet physics engine to build a simulation environment for the robotic arm to grab objects. Then, an algorithm model is established, and Markov models the capture control process. The input information of the capture decision process only includes sensor observation input (camera images), the posture of the robotic arm; the reward method uses sparse rewards, which can only be obtained when the capture is successful Positive feedback; the continuous action space at the end of the robotic arm in the Cartesian coordinate system is used as the output of the decision. Next, grabbing experiment based on DDPG and NAF network. Finally, we analyze the experimental results and see if they can be transplanted to the actual platform for experiments.