Self-supervised Distillation for Computer Vision Onboard Planetary Robots

In situ exploration of planets beyond Mars will largely depend on autonomous robotic agents for the foreseeable future. These autonomous planetary explorers need to perceive and understand their surroundings in order to make decisions that maximize science return and minimize risk. Deep learning, with its demonstrated performance on a diverse range of computer vision and image processing tasks, has become the de facto approach powering many terrestrial autonomy applications from robotic vacuum cleaners to self-driving cars. However, robotic space missions present several challenges that need to be overcome before onboard deep learning-based perception can be widely adopted, one of which is the limited computational budget on such missions. Due to power constraints and extensive flight qualification requirements (e.g., radiation tolerance), current space-qualified hardware heavily relies on legacy technologies, resulting in computational limitations that preclude the use of deep learning models for real-time robotic perception tasks such as obstacle detection and terrain segmentation. Recent research in computer vision demonstrated the benefits of network distillation in developing smaller, "mobile-oriented" versions of large, state-of-the-art vision models. This paper leverages self-supervised distillation to alleviate the impact of decreasing model size on a model's performance across two Mars image benchmarks, one on surface image classification and the other on surface terrain segmentation. Using a set of 100,000 unlabeled images taken by Curiosity and large self-supervised vision models, we distill a variety of small model architectures and evaluate their performance on the published test sets. Experimental results show that on the MSL v2.1 classification task, the best-performing student ResNet-18 model is able to achieve a model compression ratio of 3.8 while even outperforming its teacher model by 1.5%. In addition, we show that using in-domain images for distillation and increasing the dataset size for distillation has a positive effect on downstream vision tasks. Overall, results indicate that self-supervised distillation enables small models to achieve state-of-the-art performance on the benchmark datasets, supporting the feasibility of performing real-time inference using these small distilled models on next-generation flight hardware such as the High Performance Spaceflight Computer (HPSC).