Background

The development of robotics has been rapid recently in different areas, for instance, industrial, social and medical. Robot navigation is a fundamental part among robotic applications in all these domains. The task of navigating from one point to another with collision-free trajectories is referred as collision avoidance. Conventional methods exploited geometric rules or required excessive amount of real world data. Without the need of expensive real world data, this project aims to show that a robot can be trained with synthetic data only in simulation environment and navigate safely. Deliverable includes a simulation environment learning framework, two high quality simulation environments and two trained navigation policies, which serve as a decision maker to navigate the robot and avoid collision.

Autonomous applications like portable robots, unmanned aerial vehicles and autonomous cars, all need an efficient algorithm to perform collision avoidance. People had succeeded with traditional methods which were mainly based on fixed rules or robot dynamics. With the advances of machine learning (ML) techniques, attempts using ML to solve traditional robotics problems have shown impressive results.

Objective

This project aims to train a deep network to perform collision avoidance through simulation using DRL. State-or-the-art DRL algorithms will be used with variation. The resulting network will be compared by carrying out extensive experiments and evaluating the network performance.

Detail report can be found here: link
Source code: RL framework | Simulation Environment