This CPS Frontier project is establishing nanosatellite constellations as sophisticated, multi-tenant platforms for space-based cyber-physical systems applications. The work is interdisciplinary, spanning controls, ML, communications, systems, and hardware.
The initiative makes constellations autonomous and equipped to compute efficiently on orbit. On-orbit computing treats constellation-level satellite control and actuation as resource management for unique nanosatellite resources, including sensor data, bandwidth, energy, and computing. On-orbit machine learning techniques bring federated learning to the constellation, creating an autonomous orbital learning system. The initiative's new communication techniques extract maximum information from each bit communicated, combining weak signals and often avoiding communication altogether.
Research areas
Orbital edge computing
On-orbit distributed computing is a core capability of a software-defined nanosatellite constellation and a requirement to implement sophisticated machine learning and constellation-scale control. The value of a nanosatellite constellation is the ability to collect and process signals inaccessible on Earth (e.g., Earth imagery). The volume of data precludes streaming all data to Earth without incurring high-ground infrastructure costs. We will equip the constellation to process data locally, at the “orbital edge,” realizing the value of orbital deployment by adding compute capabilities with efficiency and capability. In addition to on-orbit signal processing, machine learning, and inference, computational capabilities in nanosatellites make possible sophisticated control- and motion-planning computations. Orbital edge computing must support multiple co-resident multi-tenant applications sharing inelastic resources (e.g., compute, energy, attitude control). To ensure security and isolation requires a new management layer that ensures application needs are met without interference between applications.
Pictured: A closeup of the Tartan-Artibeus-1 batteryless nanosatellite, launched January 2022.
Planning, control, and actuation
We are developing new decentralized algorithms for constellation-wide motion planning and control to coordinate maneuvers across multiple spacecraft that meet mission objectives. Imaging, communication, and solar energy collection are inherently coupled to a spacecraft’s orbital location and attitude (orientation). Cameras must be pointed at targets on Earth, antennas must be pointed at ground stations, and solar panels must be pointed at the sun. Planning and control are complex because these objectives often conflict; actuation consumes scarce energy, and the system must respect hard constraints, such as keep-out zones (e.g., not pointing the camera at the sun) and actuator torque limits. Today, humans are heavily involved in pre-planning and uplinking maneuver commands to meet mission objectives and ensure safety, leading to both high latency and poor scalability to large constellations. We are advancing the state of the art by developing control algorithms that can run directly on the constellation and can reason about the complex interactions of multiple spacecraft, energy limitations, communication links, and actuator constraints. Our goal is to make future spacecraft constellations highly autonomous and robust to individual failures.
Pictured: An image of Emily Ruppel testing the Tartan-Artibeus-1 nanosatellite.
Constellation-wide learning on orbit
Nanosatellite constellations collect rich real-time data such as high-resolution images that can be used for critical intelligent decision-making–for example, detecting a ship in the ocean. The current approach is for the satellites to send their collected data to the ground station. The computing infrastructure at the ground station is then used to train machine learning models using these data, or to perform inference on pre-trained models. The research in this center will avoid downlinking data to the ground and instead perform on-orbit machine learning training and inference. Orbital edge computing enables a constellation to autonomously perform inference and machine learning on high-rate sensor data, with minimal or no involvement from ground infrastructure. Avoiding downlinking, constellation-based training saves bandwidth, energy, and time, enabling faster and more efficient machine learning.
Pictured: A close of up of the Tartan-Artibeus-1 batteryless nanosatellite, launched January 2022.
Communications
The center will push the limits of wireless connectivity between nanosatellites and ground infrastructure. Indeed, nanosatellites provide a challenging context for wireless communication with long range, limited form factor, compute, and energy constraints. This challenge is further coupled with limited and sporadic access to satellite ground infrastructure that caters to nanosatellites. In effect, these lead to both poor communication bandwidth and availability of communication links. Our work remedies these challenges through novel designs of satellite ground infrastructure and the communications stack within the satellite. We further explore the use of machine learning and data compression techniques to co-optimize wireless communications with higher-layer applications to make the best use of available bandwidth.
Pictured: The Tartan-Artibeus-1 development team, with PhD students Brad Denby and Emily Ruppel holding the nanosatellite.