MSc research project on using a curriculum learning approach to train a robot arm with deep reinforcement learning.
Curriculum learning consists in breaking down a complex task into a sequence of subtasks of increasing difficulty. Bengio et al. showed that curriculum learning approaches can shorten training time and improve generalization for supervised learning tasks, similarly to unsupervised greedy layer-wise pre-training.
The purpose of this work is to evaluate the effectiveness of curriculum learning for robot manipulation tasks, by implementing a Deep Q-learning algorithm, following the recent success of deep Q-networks at training RL agents with human-level performance at playing many Atari games (Mnih et al.)
Main files are located in curriculumlearning/scripts/v-rep_project/ :
-
training_independent_joints.py implements the Deep Q-learning algorithm
-
robotenv.py implements an interface to communicate with V-REP simulator based on OpenAI Gym environments
-
MicoRobot.ttt is a V-REP scene file containing a model of the Mico Robot Arm