Autonomous Small Body Gravimetry via A2C Path-Planning

The success of proximity operations near small bodies relies on proper characterization of the corresponding gravitational environment; the accuracy of the gravity field model is a critical element to plan safe spacecraft trajectories and constitutes a crucial aspect for the definition of the spacecraft dynamics. Currently, flight operations required for an accurate reconstruction of the gravity field are orchestrated by ground control personnel; however, automatizing such flight control processes may yield reduced operational costs and additional mission opportunities. The problem of autonomous gravity field reconstruction can be formulated as a Partially Observable Markov Decision Process; in this framework, a spacecraft moving in an unknown gravitational environment can be modeled as an agent that autonomously implements a guidance policy to obtain accurate grav-imetric measurements. Possibly compatible with limited on-board resources, advances in flight autonomy may be sought through the exploitation of novel techniques based on Reinforcement Learning (RL) and Artificial Neural Networks (ANN). The architecture proposed in this work employs a Hopfield Neural Network (HNN) for the reconstruction of the gravity field, which is represented as a spherical harmonics expansion, assuming an Exterior Gravity Field Model. The agent’s objective is to determine a trajectory around the target body that would allow the quick and precise estimation of the spherical harmonics coefficients via HNN. The algorithm adopted is the Advantage-Actor Critic (A2C), where the agent plays the roles of the Actor; such RL algorithm exploits two networks that work in parallel aiming to maximize the return, a scalar value that renders the accuracy of reconstruction of the gravity field. In particular, this works focuses on the reconstruction of the first zonal Stokes’ coefficient C2, testing the architecture on specific case studies, as well as on more generic environments. The ANN are updated using an Adam’s algorithm for the learning process, which is driven by a reward function designed to retrieve the expansion coefficient in a quick and safe manner. Results presented in this paper show that an agent with proper training performs better than one that follows random behavior, achieving the desired accuracy more often than in a random policy simulation, in a wide pool of scenarios (different initial conditions for the same asteroid and different asteroid); in addition, gravity coefficient reconstruction performance are improved if an expert-knowledge is infused into the training process. Such results allow to assess the feasibility of the method proposed, thus defining a promising starting point for further developments.