Generating 3D Indoor Maps Autonomously Through Reinforcement Learning

In September 2017, Unity Technologies released the first open beta of the Unity Machine Learning Agents Toolkit. With this toolkit, it is possible to train agents (e.g. through reinforcement learning) to solve a specific task in a simulated environment.
The ongoing effort to automate tasks in different industries (e.g. manufacturing or logistics) could benefit from autonomous agents (e.g. robots or self-driving vehicles). It has many advantages to being able to simulate an unit (e.g. a vehicle) and the environment it acts in for training an autonomous agent. E.g. you can simulate the unit and the environment hundreds of times in parallel to speed up the training process. Additionally, you don't need to acquire any expensive hardware for building a unit.
But how complex can a task in such a simulated environment be, which we want to solve? To explore this issue, we want to train an agent using reinforcement learning that can be used to autonomously move an unit through a room. At the same time, the agent should be able to scan the room and generate a 3D map of it.

  • Goals

    The goal of this work is to generate a 3D indoor map using an autonomous agent.

    The agent must be able to pilot the unit to move through and scan the indoor space autonomously by means of reinforcement learning without any prior knowledge of the indoor geometry.

  • Tasks

    • Make yourself familiar with the Unity framework and the Unity Machine Learning Agents Toolkit (ML Agents Toolkit)
    • Reproduce some of the ML Agent Toolkit examples
    • Design a suitable reward function for the agent to generate a complete 3D indoor map
    • Design a strategy to tackle the complexity of the overall task (e.g. reduce the problem space into more manageable sub-spaces)
    • Design a unit for moving through a room and scanning it at the same time
    • Design different virtual environments (i.e. rooms) the agent should generate a 3D indoor map of
  • Technologies

    • Machine Learning, Deep Learning, Computer Vision
    • Unity framework, ml-agents library
    • C#, Python
  • Further Information

    • 30% theory, 70% software engineering
    • Master’s thesis, 1-2 students
    • Prior knowledge recommended in
      • Machine Learning, Computer Vision basics
      • Unity framework basics
      • C#, Python

Haben wir dein Interesse geweckt?

Ich möchte mich auf die Stelle als Generating 3D Indoor Maps Autonomously Through Reinforcement Learning bewerben.