Device for autonomous rocketry
A device for controlling a rocket to account for environmental uncertainties and maintain optimal mission performance. In embodiments, the device is a radiation hardened field programmable gate array with two embedded artificial intelligence programs contained in a graphics processing unit that is used for rocket reaction control by generating thrust vector commands.
1 . A computing device for a rocket's reaction control comprising at least two artificial intelligence computer programs, the computing device comprising:
a radiation hardened processor comprising at least one graphics processing unit and at least one central processing unit,
the radiation hardened processor further storing a first artificial intelligence computer program,
wherein wiring connects the radiation hardened processor to the rocket's thrust vectors further comprising a fuel injector, the fuel injector injecting fuel to one or more engines according to commands produced by the first artificial intelligence computer program,
the first artificial intelligence computer program being a simulation trained and containing a plurality of neural networks for processing real-time sensor data, including at least one convolutional neural network for processing camera data, at least one convolutional neural network for processing LiDAR data, and at least one convolutional neural network for processing GPS data,
wherein, the data processed by the plurality of neural networks are synthesized through a point cloud generation mechanism, aggregating the output of the at least one convolutional neural network for processing camera data, the at least one convolutional neural network for processing LiDAR data, and the at least one convolutional neural network for processing GPS data,
the first artificial intelligence computer program further generalizing about the rocket's trajectory environment and producing labelled visual data by processing a synthesized real-time sensor data through a sensor fusion mechanism for the purpose of automating computer vision,
a second artificial intelligence computer program comprising at least one reinforcement learning program, processing the labelled visual data and producing optimal commands for reaction control, controlling the rocket's thrust vectors to minimize error,
wherein the at least one reinforcement learning program utilizes a trained proximal policy optimization algorithm as a mechanism for policy selection, the policy selection mechanism being deployed for action selection in a continuous state space,
the second artificial intelligence program further controlling thrust vectors, by the trained proximal policy optimization algorithm selecting actions in a continuous state space, to manipulate the rocket's thrust control and steer the rocket's pitch, attitude, roll, and yaw, and
the second artificial intelligence program further comprising a deep reinforcement learning computer program optimizing control metrics corresponding to distance, time, and impact via an explicit reward function.