Description

The main problem of using a real robot to learn a task by reinforcement learning is that a lot of policies have to be tested in order to find the appropriate one, and this is really time-consuming because requires a lot of executions and has the risk to damage the robot. In order to avoid this, the robot reinforcement training can be done in simulation, where policies can be tested much faster and there is no risk to damage the robot. But this approach has a problem: what has been shown to work in simulation tends to fail in the real robot.

The work is under the scope of the following projects:

• HuMoUR: Markerless 3D human motion understanding for adaptive robot behavior (web)
• SIMBIOTS: Facilitar una introducció de la robòtica a nous processos i aplicacions dintre de la industria (web)