Title: Multi-Objective Reinforcement Learning of Stable Controllers with Multimodal Sensing  for Robotic Manipulators in Human Environments

Faculty: Barış Akgün (baakgun@ku.edu.tr)


“Manipulation” is effortless for humans. For example, we do not need to think about how to grab objects and use basic tools. Yet we can only program very specific skills that are not robust to changes for robotic manipulators. Our physical features, perceptual capabilities, and mental faculties have evolved for millions of years to give us such sensorimotor skills. It is no wonder that robotic manipulation is in its infancy despite all the effort in the last 50 years. 

However, there is hope. The development of safe and affordable robots, low-cost sensors, and most importantly the rapid advances in artificial intelligence, specifically deep reinforcement learning, shows promising potential to enable manipulators to develop sensorimotor skills for uncertain and dynamic environments.

The challenges are:

  • Working with multiple sensory modalities (e.g. from vision to touch) to perceive the environment
  • Defining or teaching reward functions at the point of operation rather than a point of development. This is related to letting the users of the robot program their robots instead of robotic developers.
  • Dealing with multiple objectives in a non-scalar manner. The standard is to have a linear combination of such objectives but this is not always satisfactory and leads to hand-tuning of combination weights. 
  • Learning stable controllers (aka policies). Since the manipulators need to interact physically with the world, stability becomes very important which is mostly overlooked in robotic RL.
  • Using human-interaction to bootstrap learning

We are looking for post-doctoral researchers to develop innovative solutions to these challenges. Interested candidates must have experience in a combination of the following:

  • Deep reinforcement learning and machine learning for control
  • Working with manipulators
  • Multi-objective optimization
  • Human-Robot Interaction
  • ROS/Python/C++ programming