PI: Danica Kragic
Learning, Interactive Autonomous Systems
To be deployed in unstructured environments, autonomous systems such as robots need the ability to learn motor skills autonomously, through continuous interaction with the environment, humans and other robots. Although classically built on rigorous control theory, mathematical and theoretical computer science methodologies, more recently data-driven learning methods, such as Deep Learning (DL) and Reinforcement Learning (RL) have been demonstrated as powerful technologies for developing AS. Still, most of the practical applications exist in carefully structured settings where i) there exists enough data to train the models, ii) one can structure the search problem efficiently, and iii) there is a computational infrastructure and/or many robots to run large scale experiments. The last point in particular is valid for only handful of labs in the world and even in their case, mostly in collaboration with a handful of companies.
This project will develop new self-supervised and meta-learning methodologies with causal reasoning for perception, control and reasoning in robotics. We aim to develop robots that achieve complex interactions with the environment, including both rigid and deformable objects, and humans. In our approach, this relates to encoding of robot motion for interaction with objects and learning of models for high-level task modeling and planning. We will build on our ongoing work in computer vision, machine learning and robotics published in leading venues. Developed self-supervised and meta-learning methodologies will, to start, address learning from small data sets and structure learning which are important open problems in ML and CV. Learning from multisensory data and learning through physical interaction will be used to address several abovementioned open problems, including interaction with deformable objects. The work will be structured along the lines of incremental and transfer learning, modeling unknown unknowns, meta-learning and causal inference.
Involved people at KTH