My goal is to achieve versatile and dexterous robots that can learn to perform new tasks quickly and be valuable assistants to humans. Model-free Reinforcement learning methods have recently shown great successes in controlling robots when trained in simulation. The transfer to the real world can be successful if the simulation is very accurate and sufficient domain randomization was performed. The latter requires a large amount of compute. Nevertheless, the resulting systems are not adaptive to unforeseen changes in the environment or the robot. How can robots learn efficiently during run-time? 
I will present recent works from my group that attempt to solve this challenge using model-learning and planning.
We study the learning of predictive models and fast planning with these models. 
I will speak about the difficulty of extracting a policy from such plans and how to use the models to perform informed exploration and to achieve safety-aware behavior. 

Download the slides for this talk.Download ( PDF, 16469.3 MB)