It is commonly assumed that the mind disposes of a repertoire of different strategies. The mechanisms underlying the selection of the proper strategy from the toolbox are still not well known. In this project we aim to better understand how strategy selection is cognitively and computationally implemented and how it can be improved.
This project will develop robotics-specific machine learning methods. The requirement for such methods follows directly from the no-free-lunch theorems (Wolpert, 1996) which prove that no machine learning method works better than random guessing when averaged over all possible problems. The only way to improve over random guessing is to restrict the problem space and incorporate prior knowledge about this problem space into the learning method.
Of course, there are machine learning methods that apply to a wide range of real world problems by incorporating fairly general priors, e.g. parsimony, smoothness, hierarchical structure, or distributed representation. However, even for solving relatively simple problems, such methods already require huge amounts of data and computation. The overall problem of robotics—learning behavior that maps a stream of high-dimensional sensory input to a stream of high-dimensional motor output from sparse feedback—is too complex to be solved by generic machine learning methods using realistic amounts of data and computation.
Many protein systems are elusive to structure analysis with established methods. This project aims to develop novel methods for protein structure determination to target this problem class of proteins. The proposed method is based on high-density cross-link/mass spectrometry (CLMS) data and custom-tailored computational algorithms to interpret them. Specifically, this project targets three critical and interdependent endeavors for advancing cross-linking for structure determination: 1) Increasing the density of CLMS data, 2) improving the distribution of CLMS data, and 3) combining high-density CLMS data with customized conformational space search algorithms.
Inspired by human grasping and manipulation capabilities, we build anthropomorphic soft robotic hands with a high degree of dexterity to enable robust interactions with the environment. We develop new sensor technologies that work with the highly compliant hands, while still providing useful sensor feedback. At the same time, we further increase the robustness of soft hands by devising control methods that reduce perceptual, model, and motion uncertainty through haptic feedback.
The "programming" of soft robots, however, remains largely an open problem. Whereas for hard robots, programming meant the specification of actuation commands, in the case of soft robots we must program control as well as "program" the robot's morphology to fully leverage the advantages afforded by soft material robotics. And, of course, these two problems interact very closely. We therefore speak of a co-design problem: we must determine control and morphology together to determine the behavior of a system.
We propose a computational principle for mapping sensory inputs to suitable actions, consisting of three building blocks: recursive estimators, interconnections, and differentiable programming. This integrated system can extract task-relevant information from the sensory input and generate suitable actions to achieve complex goals. We seek to study it as a model for different intelligent behaviors, thereby proposing it as a more general principle of intelligence.
The main obstacle to a wide-spread adoption of advanced manipulation systems in industry is their complexity, fragility, lack of strength, and difficulty of use. This project describes a path of disruptive innovation for the development of simple, compliant, yet strong, robust, and easy-to-program manipulation systems. The idea is: Soft Manipulation (SoMa).
Robots need to be able to explore and understand their physical environment. In this project we researched skills that enable robots to physically interact with their environment and explore it in informative ways.
In May 2015, our Team RBO won the Amazon Picking Challenge. This challenge addressed one of the last problems in warehouse automation: identifying and grasping objects from a cluttered warehouse shelf. Our robot was able win the competition by picking 10 out of 12 objects, outperforming 25 teams from Europe, USA and Asia.
Motion Generation is concerned with the planning and execution of motion tasks while avoiding collisions. We are especially interested in motion generation for mobile manipulators operating in the real world. Mobile manipulation takes place in unstructured environments which still poses a challenge for motion generation method. Detailed and reliable models of unstructured environments are usually not available.