The article discusses a new training scheme for distributed robot systems in open-world scenarios. It uses a teacher-student model where a ‘student’ robot can ask other ‘teacher’ robots for guidance in unfamiliar places. The training data is reconstructed from the teacher model and used for continual learning of the student model. The scheme makes minimal assumptions about the teacher model, allowing it to handle various types of open-set teachers. The paper also explores a ranking function as an instance of such generic models, using a challenging data-free recursive distillation scenario.
Publication date: 26 Dec 2023
Project Page: https://arxiv.org/abs/2312.15897v1
Paper: https://arxiv.org/pdf/2312.15897
