We propose to create a deep reinforcement Machine Learning (ML) system and development approach that supports certification for mission-critical applications through observable, verifiable architectures and functional safety methodologies, to handle the full scope of onboard, autonomous spacecraft guidance, navigation, and control (GNC). ML systems are currently being used for GNC in many autonomous systems, the biggest investments being for self-driving cars and robotics. We will investigate the feasibility of adopting certification standards based on the latest developments from the automotive industry combined with traditional aerospace certification processes. In Phase 1 we will implement a low fidelity ML-based GNC system in order to demonstrate the viability of the ML approach, whereby the trajectory plan is determined by a neural network and the control loop is executed with an Extended Kalman Filter. This demonstration will inform the drafting of requirements and specifications for a functional safety development framework.

The ability to reduce costs when exploring complex gravitational environments, (e.g. at uncharacterized asteroids), the communications bandwidth, ground system resources, and labor required to develop and verify the gravity model, trajectories, and failure modes for successively closer passes, orbits, and landings are a significant cost and schedule drivers. The ability to have satellites rendezvous in deep space, possibily for refueling.

Our ProxOps solution will allow for a significant advancement in satellite independence from Earth supervision while minimizing spacecraft burden. This AI will significantly reduce the risk and cost associated with operating spacecraft close to a small asteroid, other satellites, and other planetary bodies by utilizing Real-time mission sequencing and safe proximity operations with near-Earth objects and Satellites.