Continual Learning for Robotics: Definition, Framework, Learning\n Strategies, Opportunities and Challenges

摘要

Continual learning (CL) is a particular machine learning paradigm where the\ndata distribution and learning objective changes through time, or where all the\ntraining data and objective criteria are never available at once. The evolution\nof the learning process is modeled by a sequence of learning experiences where\nthe goal is to be able to learn new skills all along the sequence without\nforgetting what has been previously learned. Continual learning also aims at\nthe same time at optimizing the memory, the computation power and the speed\nduring the learning process.\n An important challenge for machine learning is not necessarily finding\nsolutions that work in the real world but rather finding stable algorithms that\ncan learn in real world. Hence, the ideal approach would be tackling the real\nworld in a embodied platform: an autonomous agent. Continual learning would\nthen be effective in an autonomous agent or robot, which would learn\nautonomously through time about the external world, and incrementally develop a\nset of complex skills and knowledge.\n Robotic agents have to learn to adapt and interact with their environment\nusing a continuous stream of observations. Some recent approaches aim at\ntackling continual learning for robotics, but most recent papers on continual\nlearning only experiment approaches in simulation or with static datasets.\nUnfortunately, the evaluation of those algorithms does not provide insights on\nwhether their solutions may help continual learning in the context of robotics.\nThis paper aims at reviewing the existing state of the art of continual\nlearning, summarizing existing benchmarks and metrics, and proposing a\nframework for presenting and evaluating both robotics and non robotics\napproaches in a way that makes transfer between both fields easier.\n