Katharina Muelling
Research Areas
Autonomy Infused Teleoperation
Teleoperating
a robot arm to perform fine manipulation and every day living tasks can
be hard due to challenges such as latency, intermittency, and
asymmetry in control inputs. These problems are particularly relevant
when the user is physically impaired and has to control the arm through
input devices such as brain computer interfaces, EMG, or 2D joysticks.
The combination of autonomous robot technologies and direct user control
inĀ a shared-control teleoperation framework can help to overcome the
involved challenges. By combining computer vision, user-intent inference, and human-robot arbitration we strive to create intuitive control that enables human-like manipulation behavior. The research question that we address in this project include intent recognition, motor skill learning, reinforcement learning, user adaptation, and autonomous mode switching.
Motor Skill Learning
Robotic
systems that are able to perform various tasks in human-inhabited and
unstructured environments require robust movement generation and
manipulation skills that compensate for uncertainties and disturbances
in the environment. Such systems need to autonomously adapt to a highly
dynamic environment while simultaneously accomplishing the task at hand.
I am interested in developing machine learning algorithms for learning
motor skills that can circumvent the limits of analytical engineered
solutions. A fundamental problem for the development of robot learning
methods is the necessity to achieve complex behaviors with a feasible
amount of training data. Human demonstrations can be used to initialize
robot learning approaches and reduce the learning time significantly.
Furthermore, it provides a natural way for humans to teach robots motor skills and
allows robots to acquire human-like behavior which is beneficial for
human-robot interaction.
Learning Higher-Level Behavior
A motor behavior is always directed towards achieving a specific goal. But what is the best way to
achieve this goal? While symbolic planning has been successfully applied
in many classical AI areas, they fail to scale to real robot behaviors
especially if human interactions are involved. This is due to their
limitations to model the uncertainty of actions, to address geometric
and kinematic constraints and to model human behavior. I am interested
in modeling and learning such higher-level decision processes to enable
efficient and human-like robot behavior and problem solving skills. In particular, I focus on: (1) developing a forward model to understand environmental changes caused by actions, (2) predicting if and when action can be applied, and (3) developing a hierarchical framework for task planning to facilitate the learning of complex tasks.
Social and Interactive Motion Planning
Creating autonomous and intelligent systems that are able to move
out of the factory floors into human inhabited
environments bears many challenges. While the state-of-the art in robot
motion planning has made tremendous progress in planning in
high-dimensional
and even dynamic environments, it is still hard for robots to navigate
through a crowded environment and to interact with humans in a safe and
socially acceptable manner. To enable the robot systems to work with
and close by humans they (i) need to be able to infer the intent of the
human and to integrate it in an efficient manner into the planning
process, (ii) behave in an human understandable manner, and (iii) interact
with the human in a social manner. This research project aims to shed light into questions such as: How do we adapt our movements with respect to others? Which humans do we pay attention to when adapting our movements? and How can we create socially acceptable behavior when navigating in crowds.