The main objective of B-LEARN II is to enhance robot programming and robot autonomy by introducing learning capabilities in order to close the loop between sensing and action. General aspects of interfacing robotics and machine learning are investigated for three applications: compliant actions in assembly, monitoring in assembly and machining, and the navigation of mobile systems. Based on these studies, the general requirements for the application of machine learning in robotics are defined. Appropriate enhancements are made to robot system architectures, leading to robot systems with learning capabilities.
Robust subsymbolic and symbolic learning strategies are being studied and developed that should prove useful in enhancing robot performance, reliability and flexibility as well as in supporting the acquisition of robot skills. These techniques are applied to real assembly, machining and navigation tasks, using a framework of enhanced robot control architectures supporting the integration of machine learning algorithms.
The work has focused on the definition of a common scenario and a sound experimental methodology. Moreover, the requirements for interfacing machine learning and robotics have been investigated on the base of given case studies.
The main results so far are:
a general architecture supporting the application of machine learning to robotics;
architectures supporting the application of machine learning to robotics which are derived from the general one but take the special requirements of the case studies into account;
a classification of the learning tasks which are to be solved within the several case studies and an experimental methodology which should lead to this solution;
exchange of knowledge between machine learning researches and robotics researchers, leading to a better understanding of the common problems, which is crucial for the success of the project;
development of experimental setups;
first experiments with several learning techniques on the base of real world sensory data, recorded from test beds.
APPROACH AND METHODS
The difficulty of closing the loop between sensing and action by programming appropriate control strategies is one of the major cost factors in robot applications. Moreover, it limits the applicability of robots to non-structured domains. Following the hierarchical structure of the target systems (robots, mobile platforms, and machining devices), symbolic and subsymbolic learning strategies and their combination (hybrid learning systems) are studied. Additionally, robot system architectures are studied, given the requirements of machine learning methods. Certain enhancements due to the aspect of interfacing are added to both components.
On the lowest level, the main problem is that of feature selection and feature extraction. Subsymbolic methods are envisaged in order to automatically refine this process.
On the second level, a conceptual structure representing the robot's internal model of the world must be created. Here symbolic learning techniques are used. The highest level deals with the problem-solving activity in order to reach the main goal of the robot, which might be reaching a given destination, assembling a component, or enhancing the current skills (experimentation). On this level, learning heuristics in problem solving is the typical learning task.
With respect to the aspect of knowledge exchange between the control level and the problem solving/planning level, methods to use low-level robot skills most efficiently in planning and to use high-level knowledge to enhance control skills, eg by experimentation, are developed.
The result of this project will be a set of case studies, demonstrating how machine learning can be applied to robotics, and giving a detailed description of requirements for certain applications.
The cross-fertilisation of the two fields of machine learning and robotics has a deep impact on both sciences. The requirements of machine learning have already lead to an evolution of the robot architectures that are used in B-LEARN II. The control and programming techniques nowadays used for robots can also be expected to change significantly.
Machine learning has until now been applied only very little to real-world problems. It is now facing real robotics task, and the existing methodologies are already and will have to be compared in the light of the concrete tasks and the concrete results that can be achieved. Results and hints from B-LEARN II should be able to influence future research in machine learning. In order to emphasise the benefits arising from the project's synergetic nature and to encourage further cooperation, a workshop will be organised after the first year.
The final results will be demonstrated and explained to the industry at a workshop to be organised at the end of the project.
2825 Monte Da Caparica