Real-time Hazard Detection via Deep Learning

HazNet is a robust hazard detection solution that leverages deep learning and hardware acceleration to achieve mission-speed performance on path-to-flight hardware. The HazNet solution seeks to maximize data use while maintaining flexibility by leveraging the independent strengths of LiDAR and camera data to produce a single hazard map. Flexibility is maintained by using two independent convolutional neural networks for computation, one for LiDAR data and one for image data, which are combined into an existing hazard map to improve knowledge, resolve unknown regions, and increase hazard map resolution. This method de-risks the transition from traditional hazard detection to deep learning-based algorithms by leveraging well-proven, rather than experimental models to identify hazards. It improves upon these traditional methods by acting on the strengths of complimentary sensors, enabled by hardware acceleration. Astrobotic proposes the development of a prototype sensor package for HazNet, while further advancing developed hazard detection models and techniques. This Phase II effort will entail five major efforts: working in collaboration with NASA and Astrobotic stakeholders to develop a reference mission and associated requirements; advancement of models developed in the Phase I effort to incorporate uncertainty and combine hazard map outputs; development and testing of custom sensor package;nbsp;demonstration in a series of relevant simulations;nbsp;and a final technology demonstration across a descent-like scenario in a lunar-relevant environment.nbsp;