Machine Learning

TODO: How to put them on one line?

While purely engineered solutions can achieve impressive results in robotics, for many practical applications the required models (for policy, dynamics, vision, etc.) are too complex to specify them by hand. Hence, data-driven methods that train such models using machine learning are often key components in our systems that we are actively developing. In broad terms, our research can be grouped into three categories: Bayesian Methods, Reinforcement Learning, and Imitation Learning & Inverse RL.

Bayesian Methods

TODO: Is it better to put researchers like this?

Bayesian methods provide a principled way to deal with the uncertainties at hand in the robot's environment. For example, we develop algorithms for Bayesian optimization (e.g. for optimizing robot controllers or risk-sensitive exploration), for variational inference (e.g. for approximate Bayesian inference or trajectory planning) or for control-as-inference.

Approximate Inference

TODO: Hany, Oleg, Joe

Bayesian Optimization

TODO

Active Learning

TODO: Boris, Tim

Reinforcement Learning

In order to apply reinforcement learning to real robots, we tackle fundamental research problems related to the efficiency of the RL algorithm, the safety during exploration, and the robustness of the learned policy. In particular, we address problems related to model-based and off-policy reinforcement learning, policy gradient estimation, and curriculum, transfer, and lifelong learning. Furthermore, several groundbreaking algorithms, such as reward-weighted regression, the natural actor-critic, relative entropy policy search and MORE originate from our group.

Curriculum Learning

TODO: Pascal

Gradient Estimation

TODO: Samuele, Joao

Model-Based RL

TODO: Daniel, Georgia

Imitation Learning and Inverse Reinforcement Learning

As reward functions usually need to be specified by experts, who are not always available, we also investigate methods that enable non-expert users to program robots my means of demonstrations. Furthermore, such imitation learning methods enable the robot to solve a given task in similar style compared to a human, which makes the robot's behavior more predictable and, thus, safer for human-robot interactions. In addition to imitation learning methods, which directly aim to infer the policy from the human demonstrations, we also develop methods for inverse reinforcement learning, which aim to infer a reward function as a concise representation of the demonstrated task. Such reward function can enable the robot to better react to changes in the environment, or to better predict the human behavior.

TODO: Oleg, Davide, Julen