Evaluating the Reliability of Vision Transformers for Space Robotics Applications

摘要

The deployment of autonomous robots in safety-critical applications, such as planet exploration, demands advanced perception and motion capabilities. Deep Learning (DL) has shown promise in addressing perception tasks, but its resource-intensive nature limits deployment on resource-constrained devices like autonomous robots. Tiny Machine Learning (TinyML) emerges as a solution, leveraging low-power and low-cost devices to reduce computational demands, potentially allowing the adoption of DL also in power-constraints space applications. However, before being adopted in safety-critical applications, the reliability of DL models to transient errors needs to be evaluated. This paper investigates the radiation reliability of transformer models deployed on a Coral Edge TPU. The reliability of six transformer models and dedicated micro-benchmarks is evaluated, which provides insights into fault propagation and their impact on model accuracy. The study encompasses over 32 hours of accelerated neutron irradiation, equivalent to more than 12.5 million years of natural neutron exposure, offering a comprehensive assessment of transformer reliability in safety-critical applications.