Joao Moura

Research project title: 
Learning manipulation tasks by demonstration
Principal goal for project: 
Formulate a method for learning the unconstrained policy of a constrained manipulation task for a torque controlled robot.
Research project: 

We are now able to efficiently learn an unconstrained policy of a constrained manipulation task, such as wiping a surface, for a velocity controlled robot. In principle this unconstrained policy can be a generic policy encoded as linear combination of local models. However, due to the high dimensionality of the space of the unconstrained policy - equal to the number of DoF of the robot - and to the little amount of data obtained in a few demonstrations, learning a generic unconstrained policy becomes unfeasible, and the learned policies tend to diverge from the desired goal. In the case of a velocity controlled robot we can use the Jacobian of the robot and the knowledge of the unconstrained task to choose some specific local models that allow learning and reproducing the demonstrated task.
When tackling the dynamic case - a torque controlled robot - it becomes difficult to specify appropriate local models for the task. Moreover, the dimensionality of the unconstrained policy increases because the state is now composed by both configuration position and velocity.
I am currently investigating the application of DMPs for learning this unconstrained policy. This will allow us to use a generic set of local models while guaranteeing that the learned motion does not simply diverge from the goal.

Publications

  • João Moura, Mustafa Suphi Erden. Formulation of a control and path planning approach for a cab front cleaning robot . In Procedia CIRP, volume 59, pages 67–71, 5th International Conference on Through-life Engineering Services (TESConf), 2017, DOI: 10.1016/j.procir.2016.09.024.
  • Leopoldo Armesto, João Moura, Vladimir Ivan, Antonio Salas, and Sethu Vijayakumar. Learning constrained generalizable policies by demonstration. In Robotics: Science and Systems XIII (RSS), 2017, DOI: 10.15607/RSS.2017.XIII.036.
  • Leopoldo Armesto, João Moura, Vladimir Ivan, Mustafa Suphi Erden, Antonio Salas, and Sethu Vijayakumar. Constraint-aware Learning of Policies by Demonstration. In International Journal of Robotics Research (IJRR), 2018. (to appear)
  • João Moura, William Mccoll, Gerard Taykaldiranian, Tetsuo Tomiyama, Mustafa Suphi Erden. Automating Train Cab Front Cleaning with a Robot Manipulator. In IEEE Robotics and Automation Letters, 2018. (to appear and to be presented at the 14th IEEE International Conference on Automation Science and Engineering - CASE)
About me: 

I hold an Integrated Master's degree in Mechanical Engineering from University of Aveiro, Portugal.
I worked for a year at Bosch Thermotecnology, Aveiro, Portugal. As a developer engineering I helped develop the control system of a gas water heater appliance and supported its industrialisation.
I worked for a year and half in the CROB (Centre for Robotics and Intelligent Systems) at INESC TEC, Porto, Portugal. As a researcher I provided technical support to ongoing projects such as the European project - ICARUS.
I was a teaching assistant at Aveiro University for Servomechanisms and Industrial Robotics master courses.
My current research interests are manipulation, robot control, and supervised and reinforcement learning.