Abstract
Autonomous robots operating in uncertain and dynamic environments require visual perception models that are both robust and efficient. Knowledge distillation provides a practical way to deploy compact student models on resource-constrained robotic platforms, but conventional distillation usually assumes that teacher predictions are uniformly reliable. In realistic visual scenarios involving sensor noise, appearance ambiguity, occlusion, and distribution shifts, teacher models may produce biased or incorrect predictions, and directly transferring such supervision can lead to negative transfer. To address this problem, this paper proposes a reliability-aware knowledge distillation framework with self-evolving supervision for robust robotic visual perception. The proposed method first separates reliable and unreliable teacher predictions according to their consistency with ground-truth labels. Reliable predictions are preserved as stable supervision, while unreliable predictions are corrected through ground-truth-constrained adaptive fusion. A Self-Evolving Knowledge Bank is further introduced to progressively refine corrected supervision through temporal accumulation and student-guided anchoring. Experiments on CIFAR-10, CIFAR-100, and Tiny-ImageNet with convolutional networks, vision transformers, and sequence-based visual models demonstrate consistent improvements over representative distillation methods. These results show that reliability modeling and progressive supervision evolution can improve compact visual models when teacher supervision is unreliable. In the present study, robustness primarily refers to resistance to negative transfer caused by erroneous teacher predictions, rather than comprehensive robustness to environmental disturbances such as sensor corruption, occlusion, or domain shift. The proposed framework therefore provides a promising training strategy for future integration into resource-constrained robotic perception pipelines.