Evaluation of combined Few-short and continual learning approaches for Robotic perception

State-of-the-art robotic learning frameworks struggle in typical real-world scenarios like dynamically changing or unseen environments that require robots to adapt quickly (few-shot and meta-learning). Furthermore, real-world settings require robots to continually learn without forgetting what they have learned before (continual learning). Few-shot learning and continual learning currently belong to the rapidly evolving research areas. How to optimally tailor few-show continual learning strategies to enhance robotic autonomy belongs to the open research questions. This master thesis aims to assess how to optimally fuse state-of-the-art and/or new few-shot and continual learning approaches applied to the problem of object classification using suitable few-shot continual learning datasets. The performance will be evaluated and compared by formulating suitable evaluation metrics addressing aspects like catastrophic forgetting, classification accuracy, required sample sizes, and the question, of whether extensive pretraining is required or not.