A remaining major challenge with autonomous systems is the handling of situations that the system cannot handle on its own. Up to now, this transfer of control has been considered in isolation. An important step towards long-term autonomous systems and the aim of the project is to develop systems that can learn from past situations in order to improve their behaviour in future, similar situations. As a result, the transfer of control to humans can become more efficient, and in some cases even obsolete, which is also a core requirement of the industry.
CAMELOT is in this sense a follow-up project of TRACTAT and builds on its results by looking at the task from the perspective of self-learning systems and multimodal human-machine interaction. Machine learning models support the system to recognize and classify situations. The models are multiply adaptive; they can be improved by passive observation and by actively being teached by user to deal with new situations. Multimodality plays a role on the one hand as a source for the recognition of user behaviour in response to the system, and on the other hand for the natural communication between system and user in case of a transfer of control. By using new methods in combining symbolic and sub-symbolic learning, not only is explainability and extensibility ensured, but the overall recognition performance is also improved compared to the state of the art.