RoboNav - Autonomous robot navigation using reinforcement Learning

Course : Reinforcement Learning

Problem Statement : To train an agent to autonomously navigate in an indoor dynamic environment for reaching a desired goal

Application : Autonomous robot navigation for indoor service robots

Approach :

• Trained the robot agent using reinforcement learning algorithms of Deep Q Network and Deep Deterministic Policy Gradient in ROS to compare the effectiveness of the two & benchmarked them with the move_base package of ROS Navigation Stack.

• The input to the agent is a laser range data around the robot and its current position & orientation which on passing through a deep network gives most feasible robot action (go straight , take a slight/steep left, take a slight/steep right).

• Achieved 70% success rate for DQN & 90% success rate for DDPG.