2019.B.2.5. Neural Attitude Control: Nanosatellite attitude control with Deep Reinforcement Learning

Author(s)

Andrea Cini (1)
Mhamed Matteo El Hariry (1)
Alessandro Balossino (1)

Argotec srl, Italy

Session

B.2

Keywords

Autonomy, AI, attitude control, reinforcement learning

Abstract

Nanosatellites benefit from reduced production cost, lower design complexity and ease of launch with respect to larger spacecrafts, therefore they are a key technology for low-cost missions. However, they still lack the autonomy required to operate without supervision from the Ground Segment. This is particularly true in the case of off-nominal events leading to sensor and actuator failures. Allowing nanosatellites to independently handle their attitude in a robust way, even in extreme conditions, is an enabling factor towards full nanosatellite autonomy: for this reason, we propose a control policy based on Deep Reinforcement Learning, which can detect and react to actuation failures, reaching its target attitude. The control signals are directly mapped from the output of a Neural Network which is trained in a simulated environment, having access only to the current orientation quaternion and angular rates of the satellite. The proposed approach has the additional benefit of enabling the designer to specify the desired control properties by simply defining the reward structure of the problem. The artificial agent does not need any prior knowledge about the problem, it directly learns how to maximize the chosen reward signal through experience. In this paper we present our method for nanosatellite attitude control, discussing its main advantages and drawbacks, presenting the roadmap for the in-orbit validation of the system on a 6U satellite that will be launched in the frame of an international mission. Finally, we show experimental results obtained from simulations on relevant nanosatellite hardware.