top of page
Department of Electrical Engineering
Control Robotics and Machine Learning Lab

Active Projects

Learning Skills to Play RoboCup Soccer from Scratch using DQN
  • Place an agent on a soccer field in RoboCup. 

  • Train agent to approach a soccer ball and score past a keeper

  • Modify pre-existing algorithm to learn this behaviour

Automatic Skill Learning for Lifelong Learning in Minecraft
  • Train an agent in Minecraft to learn skills automatically to solve a given task

  • When the agent is given a new task, refine the skills to solve this task

 

  • Build on a pre-existing Minecraft Deep Learning work that has been published from our lab in AAAI 2017!

The OpenAI Universe Challenge
  • Find a domain in OpenAI Universe that has never been solved before

  • Using state-of-the-art Deep RL algorithms to solve the domain

Successor Features for Minecraft
  • Learn to solve a set of tasks in a Minecraft domain

 

  • Investigate methods for performing transfer between the tasks to solve the tasks faster or generalize to new unseen tasks.  

Advanced optimization methods for DQN
 
  • Develop and implement various optimization techniques such as SVRG in the DQN framework to improve upon current state of the art results.

Advanced optimization methods for LSTMs
 
  • Develop and implement various optimization techniques similar (but not limited) to 'Shallow Updates for Deep Reinforcement Learning' on LSTM based architectures.

Action eliminating DQN: exploration via confidence bounds
  • Build upon the recent developments by Google DeepMind.

  • Implement well established models with convergence guarantees in a Deep Reinforcement Learning setup.

Besting ATARI via human-robot cooperation
 
  • Develop a human-robot interactive interface. In this interface, you will be able to replay episodes played by the agent and to manually fix it's errors.

bottom of page