Soft Hands represent a new direction in robot hand design, departing from the traditional approach that prioritizes precise models and planning of contact points. Our focus is on enhancing robustness and safety through the use of soft materials and flexible mechanics. This softness affords us the opportunity to leverage contact with the environment and employ it in efficient grasping and manipulation strategies. Our research aims to investigate various aspects of soft hand design, including the integration of sensors and feedback control methods for grasping and in-hand manipulation that utilize the hands' morphological properties.

Moreover, we view soft hands as a valuable computational resource that can support robust manipulation behavior generation. This highlights the need for a balance between the software that controls the hand and the morphological computation provided by the body, leading to the important problem of co-designing both the body and control. This approach recognizes that the hand's morphology can be harnessed to achieve efficient and effective manipulation, and that the software should take advantage of this computation.

The core idea of interactive perception is that action and perception are intertwined: they cannot be separated and must be approached together. We do not just need to perceive the world so that we can act in it, but it is often also crucial to act so that we can perceive better.

Examples for interactive perception in humans are numerous. Often we move our heads and eyes to see, but we also use our hands to search for objects, move visual obstructions out of the way or to explore haptic and dynamic properties, such as the weight of objects. Many of these perception tasks are challenging or even impossible if we were purely passive observers. Thus, we need to interact with the world to perceive it reliably.

By improving perception through action, robots can better understand their environment and perform tasks more efficiently and effectively. Robots can become more autonomous, when they actively seek the information they need to act in the world. This also benefits robustness, as robots can gather more reliable information when they move and use their sensors with purpose, for example to look at an object from a different viewpoint.

Much of the research in our lab revolves around the perception, estimation and manipulation of articulated objects. Our robots manipulate drawers, doors, and also more complicated objects such as mechanical puzzles. We aim to improve these robot's behavior not only by developing novel robotics techniques, but also by taking inspiration from biological behavior in humans and other animals.

The pursuit of robot learning aims to create robots that can adapt and improve their performance over time through the use of machine learning algorithms. These algorithms rely on data to learn from and train models that can make predictions and take actions in real-world scenarios. However, without taking the embodiment of the machine learning system into account, these models may struggle to perform effectively in real-world situations or generalize to new environments.

Embodiment refers to the physical form and structure of the machine learning system and its interaction with the environment. In the case of robots, this encompasses factors such as the robot's body shape, movement abilities, and sensory input. These elements can greatly affect a robot's ability to perform tasks and interact with its surroundings. To optimize machine learning for the field of robotics, it is crucial to consider both learning and embodiment together. We are approaching this by discovering robotics-specific prior knowledge and incorporating it into our learning algorithms. As an example, our research has shown that robots can efficiently learn versatile manipulation skills from just a single human demonstration by utilizing the benefits of embodiment to generate complementary information. Additionally, incorporating knowledge of physical laws helps us to learn state representations from data more efficiently, as robots interact with the physical world through their bodies.

Robotics is a discipline that aims to build and execute behavior. As such, it can learn a lot from biology, another research area that is also studying behavior, but in this case the behavior of animals. Animals, including humans, show generalization capabilities, problem-solving skills, and adaptivity that outshine current robots in many ways. As animals often behave with remarkable efficiency and robustness, it makes sense to look for solutions to robotic problems not just in engineering sciences, but also in biology!

Several of our past and ongoing projects aim to connect the worlds of robotics and biology in one way or another. We performed human grasping studies where we learned that robots should exploit contact with the environment when they grasp objects, and that it highly important to build robot hands compliantly. We analyzed how humans use heuristics for ball catching and how the regularities they exploit could also be used by autonomous machines. Furthermore, we try to understand how cockatoos are able to solve complex mechanical puzzles, and we aim to use this understanding to build robots that can explore and use human-made mechanical objects, such as drawers and cupboards. Related, we also research how humans select different strategies from a "mental toolbox" to explore mechanical puzzles.

These are just some examples of how we try to connect biology and robotics. While we perform this research, we often realize that it is not sufficient to employ well known and established research methodologies, but that we also need to develop the research methodology itself.

Proteins are vital to the proper functioning of our body, with their three-dimensional shape playing a crucial role in interactions with other molecules. However, observing protein motion directly is difficult due to their size and complexity. The Robotics and Biology lab is developing a biologically accurate simulator for protein motion, a computational microscope, to focus on areas of the protein that exhibit function-relevant motions.

Accurately predicting the three-dimensional structure of proteins is another challenge in structural biology that would lead to scientific advances and help find cures for many diseases. A proposed computational framework for protein structure prediction combines robotics and machine learning algorithms with molecular biology techniques to guide conformation space search using target-specific information. This approach tailors conformation space exploration to the particular characteristics of the target, resulting in highly accurate and efficient structure prediction.