Sensorimotor verilerinin verimli öğrenilmesi

Learning from Demonstration (LfD) is an effective way of equipping robots with dexterous skills, which involves learning from the temporal sensorimotor data generated by a demonstrator, typically a human. In spite of the common use of LfD for robot learning, there are still challenges to be addressed and improvements to be made. Towards this end, in this thesis, we propose a novel resource-efficient Learning from Demonstration (LfD) method and present how such LfD methods can be used for efficient reward function estimation given a set of sensorimotor data for completing a given task execution. To be concrete, we develop a divide-and-conquer approach applicable to encoder-decoder type deep neural networks and apply it to combine multiple Conditional Neural Movement Primitives (CNMP) for representing different segments of a sensorimotor trajectory that needs to be learned. By assigning dedicated network components to specific trajectory segments and coupling them in the latent space, the developed system, Coupled Conditional Neural Movement Primitive (C-CNMP) model, improves resource efficiency and achieves a significant gain in prediction accuracy. The learning and generalization performance of the proposed resource-efficient C-CNMP model has been measured with experiments conducted using synthetic data and real-world robotic movement datasets. Throughout systematic experiments with equally sized two-segmented C-CNMP models, the strengths and limitations of the proposed C-CNMP model have been analyzed. Additional investigations are conducted to further explore the characteristics of C-CNMP in multi-segment settings which can be used to represent complex movements in a structured fashion. To this end, the impact of varying the number of segments is assessed by showing the trade-off between segmentation granularity and learning performance. Another dimension explored is the segment size, where the segment count is kept fixed and the segmentation boundaries are changed. This complements the systematic analysis conducted with equally sized segments to evaluate the overall performance. Overall, the segmentation-related analyses indicate that the search for optimal segmentation can be worthwhile for both performance and resource economy. During C-CNMP research, several challenges have been encountered and addressed, including formalization of an appropriate loss function and determining the optimal number of segments. We systematically analyzed these issues and proposed solutions through empirical studies, refining the segmentation process to improve generalization. In addition to developing the C-CNMP model and investigating its properties, we have shown the application of CNMP-type methods in Inverse Reinforcement Learning (IRL) to uncover the reward function that is responsible for generating an observed sensorimotor data during a task execution. In brief, we develop a novel IRL mechanism with high computational efficiency during reward function inference. This is achieved by associating reward parameters with trajectory distributions that are represented by CNMP-type models that facilitate the construction of effective similarity metrics for reward function estimation. The proposed IRL pipeline enables fast and efficient reward inference as it does not require an inner reinforcement learning loop for reward inference. Overall, the thesis contributes to the state-of-the-art in LfD with C-CNMP that can learn and generate sensorimotor data efficiently, as well as providing a practical inverse reinforcement learning system that can be used to uncover the optimality principles underlying a task execution given a set of observed sensorimotor data.