This repository serves as the collaboration of practical project computational neuro engineering NST, TUM.
The main goal of this task is to make the robot arm behave properly to reach a pregiven position by applying spiking neural network based deep reinforcement learning algorithm.
However, some high level goal could also be possible after the achievment e.g. detect and push objects to a given region or more generally take the images as the state (spiking CNN).
The basic ideas come from the paper published by Google DeepMind Playing Atari with Deep Reinforcement Learning (https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf) and Continuous Control with deep reinforcement learning(https://arxiv.org/pdf/1509.02971.pdf).
More specificly, our approach is to apply spiking neural network as the action-value approximator in a standard Q-learning framework. This idea will be firstly verified in the mathematical emulator and transplanted to the real robot afterwards.
git clone [email protected]:356255531/SpikingDeepRLControl.git cd SpikingDeepRLControl make emulator (robot_arm) # if you want run it an real robots
Zhiwei Han, Bo Huang, Meng Wang.