We believe that increased autonomy and intelligent computing via multiagent AI will "further reduce the burden and cost of the ground segment and mission operations in CubeSats" through "onboard data processing, autonomous systems, and navigation" as quoted in the 2016 Study Report Achieving Science with CubeSats: Thinking Inside the Box. The Decadal Survey identified scientific observation and sample return as a priority. We propose our new TeamAstro as a multiagent AI for spacecraft team coordination and sensing for science return. TeamAstro contributes to increased capabilities on each spacecraft and as a team of spacecraft swarm, increased autonomy in decisions and local actions, and thus minimized communication and control bottleneck with Earth. Our innovation TeamAstro constitutes (1) a derivative of a single DeepMind’s MuZero for spacecraft application with stringent space domain constraints of orbits, fuel, etc.; (2) enabling heterogeneous single MuZero agents to work together as a team, constituting our new TeamAstro, via attention mechanism and coordination graph learning, while keeping MuZero’s rules-discovery and Monte-Carlo Tree Search (MCTS) value search, and (3) a design reference Mars mission of TeamAstro for coordinated remote sensing between multiple spacecraft. Based on the goal configuration of the spacecraft swarm and the baseline trajectory in training datasets, TeamAstro will design detailed trajectories considering the observation/control uncertainties and constraints to reconfigure the original swarm configuration to a new one. In this case, the actions would be the thrust firing sequence of each swarm spacecraft at each time, the states would be the spacecraft position/attitude and their velocities, and the reward would be defined based on the achievement of the goal configuration. The training can be performed on the ground beforehand so that the new-optimal policy can be executed on board with a limited computational resource.
Our innovated TeamAstro multistep lookahead rules-discovery multiagent algorithm will be of interest to commercial SmallSat companies as it enables more autonomy and decision making on board for both payload operations, sensor operations, energy/fuel management, and self-preservation. We plan to team up with DigitalGlobe or Teledyne Brown Engineering to improve on on-board autonomy and decision making to more efficiently return observations valuable to NASA. Together we will approach NASA via Commercial SmallSat Data Acquisition (CSDA) Program.
There is a growing need for satellite swarms for regional-coverage satellite constellations, which require a smaller number of satellites and reduced system costs. Target applications are remote sensing and docking, local high-speed Internet access constellations, localized high-bandwidth disaster response coverage, and localized bandwidth support for logistics.