Training Self-localization Models for Unseen Unfamiliar Places via Teacher-to-Student Data-Free Knowledge Transfer

TL;DR

The paper tackles open-world self-localization under no annotated data by proposing data-free teacher-to-student knowledge transfer (DFKT), where a student interacts with encountered teachers that are self-localization systems to generate pseudo-training data for continual learning. It introduces four schemes—ReplayCL (data-rich baseline), Reciprocal Rank (RR) sampling, Entropy-based sampling, and Mixup—to balance self-localization performance against communication cost $T$, and demonstrates stability of improvements in a recursive KD setting. Using a scene-graph embedding (GCN-based) evaluated on the NCLT dataset, the study shows that Entropy and Mixup are particularly effective at low KT budgets, while Replay provides the strongest performance when data access is allowed, and RR offers a surprisingly strong data-free alternative. Overall, the work enables robust self-localization in unknown environments with diverse and privacy-preserving teacher models, expanding practical applicability of open-world distributed robot localization.

Abstract

A typical assumption in state-of-the-art self-localization models is that an annotated training dataset is available in the target workspace. However, this does not always hold when a robot travels in a general open-world. This study introduces a novel training scheme for open-world distributed robot systems. In our scheme, a robot ("student") can ask the other robots it meets at unfamiliar places ("teachers") for guidance. Specifically, a pseudo-training dataset is reconstructed from the teacher model and thereafter used for continual learning of the student model. Unlike typical knowledge transfer schemes, our scheme introduces only minimal assumptions on the teacher model, such that it can handle various types of open-set teachers, including uncooperative, untrainable (e.g., image retrieval engines), and blackbox teachers (i.e., data privacy). Rather than relying on the availability of private data of teachers as in existing methods, we propose to exploit an assumption that holds universally in self-localization tasks: "The teacher model is a self-localization system" and to reuse the self-localization system of a teacher as a sole accessible communication channel. We particularly focus on designing an excellent student/questioner whose interactions with teachers can yield effective question-and-answer sequences that can be used as pseudo-training datasets for the student self-localization model. When applied to a generic recursive knowledge distillation scenario, our approach exhibited stable and consistent performance improvement.

Figures (3)

• Figure 1: Our goal is to design an excellent questioner (i.e., student) that can obtain a sequence of question $x_t$ and answer $y_t$ pairs via interactions with a blackbox teacher $y=f(x)$ such that the obtaiend samples $\{(x_t, y_t)\}_{t=1}^T$ can then be used for supervised learning or distillation of the student self-localization model. This is relevant to the model inversion problem, except that we are targeting generic self-localization models.
• Figure 2: A typical communication protocol: (1) The student robot sends the questioner (program code) to the teacher robot. (2) The questioner repeats the question and answer session with the teacher. (3) The student robot receives the selected sample via the KT channel. In this work, we aim to obtain a small number of good samples.
• Figure 3: Top-1 accuracy performance vs. KT cost ($T$).