By Lucy Wang and Rama Afullo
Two UC Berkeley MEng Capstone projects, Project 40: Small Language Models as Edge Orchestrators for Satellite Analysis and Project 126: Hardware in the Loop Testbed for Low Earth Orbit Satellite Communication and System Simulation, are exploring how satellites can operate more intelligently and independently in space.
Traditionally, satellites collect data and send it back to Earth, where humans analyze it. As satellite constellations grow larger and more complex, this centralized approach becomes slower and less efficient. These two projects aim to address that challenge by combining onboard artificial intelligence with realistic testing of satellite communication systems.
Developed in collaboration with industry partner Satlyt, the projects examine how satellites might analyze data, detect issues, and coordinate with one another directly in orbit.
At Satlyt, Founder and CEO Rama Afullo provided strategic direction, with technical oversight from CTO Nelson Kigen Psenjen and UC Berkeley Faculty Advisor Professor Panayiotis Papadopoulos. Day-to-day execution and student coordination were led by Junn Hope Wangari for the Hardware in the Loop team and Leina Meoli for the Small Language Model team.
Rethinking Intelligence in Orbit
Project 40, led by Nathan McNaughton, Youssef Miled, and Tian Herng Tan, focuses on enabling reliable decision-making directly on satellite hardware.
The team designed a hybrid system built around a Small Language Model. A Small Language Model is a compact artificial intelligence system that can process and generate text, similar to larger systems such as ChatGPT, but designed to run on devices with limited power and memory. Rather than performing every task itself, the model acts as an orchestrator. It interprets incoming requests and selects the most appropriate software tool to handle the task, such as detecting anomalies in telemetry data or summarizing system logs.
“We were drawn to this because satellites are increasingly becoming computing platforms, not just data relays,” said the SLM team. “As constellations scale, they cannot rely on constant ground intervention. The system needs to reason locally.”
In space, computing resources are extremely limited. Satellites operate within tight power budgets and have restricted communication bandwidth to Earth. Because of these constraints, the team benchmarked multiple candidate models, evaluated their memory usage and response times, and designed a tool-calling architecture that could operate fully offline.
The final system was deployed on NVIDIA Jetson edge hardware, demonstrating that the full pipeline could run locally without cloud dependency. Instead of transmitting large volumes of raw telemetry data back to Earth, the satellite calls specialized analysis tools to detect anomalies and generate concise summaries of system behavior. This reduces the amount of data that must be downlinked while preserving the accuracy of important event detection.
“Instead of selecting models based on reputation, we relied on structured metrics and testing,” the SLM team explained. “Tool calling reliability, latency, and consistency were critical to making the system work in constrained environments.”
Simulating a Satellite Network on Earth
Project 126, led by Shreya Sinha, Sahil Singh, Lanny Tseng, and Mike Lin, focuses on testing how intelligent satellite communication systems behave under realistic conditions.
The team built a Hardware in the Loop testbed. A Hardware in the Loop system combines real physical hardware with simulated environments. In this case, embedded compute nodes represent satellites and run on robotic platforms, while software simulates orbital movement and network conditions. The system recreates how satellites move relative to one another and how communication links form and change over time.
“This platform allows us to test routing logic, latency behavior, and communication disruptions under realistic conditions,” said the HIL team. “Simulation alone can hide bottlenecks. Hardware reveals them.”
Some communication links are implemented over real wireless connections, which naturally introduce variability, such as latency fluctuations and packet loss. Other parts of the constellation are simulated in software to maintain scalability. This hybrid approach allows researchers to observe how routing decisions perform under both modeled and physical constraints.
Integrating robotics, networking, visualization, and machine learning routing into a single coordinated system proved to be one of the most challenging aspects of the project. Physical systems introduce variability and unpredictability that purely digital simulations do not capture.
Bringing It All Together
Both projects build on established research in edge artificial intelligence, anomaly detection, and satellite network modeling. Project 40 draws from advances in compact language models and tool-based AI systems. Project 126 draws inspiration from prior satellite testbeds but extends them through a hybrid synthetic and physical constellation design.
What makes these projects distinctive is how they connect intelligence and validation into a unified framework. Instead of building artificial intelligence models or simulation platforms independently, the teams integrated onboard decision-making with a physically grounded communication testing environment.
“The goal is not just to build smarter satellites,” said Afullo. “It is to build systems that can think, communicate, and validate their own decisions under real constraints. That is what autonomy in space actually requires.”
Under the coordination of Junn Hope Wangari and Leina Meoli, the two teams aligned their systems toward a shared integration layer in which the Hardware in the Loop environment tests Small Language Model intelligence in a controlled but dynamic setting.
The result is a glimpse of a more responsive and resilient space infrastructure. Satellites do not need massive cloud-scale models to become more capable. Carefully designed small models, paired with structured tools and validated through hardware-grounded experimentation, can operate under tight constraints and still deliver meaningful autonomy.
As satellite constellations grow and demand for real-time, space-based data increases, systems like these suggest a future in which satellites can analyze information locally, share processing responsibilities across a network, and respond more quickly to changing conditions in orbit.