Open-Source AI Stack for CubeSat Autonomous Navigation

Open-Source AI Stack for CubeSat Autonomous Navigation presents a complete framework for deploying deep reinforcement learning on resource-constrained spacecraft. The paper details a Hierarchical TD3 architecture trained on consumer GPUs (NVIDIA 4070 Super) and optimized for flight hardware with only 256MB RAM. It covers the full pipeline—from TensorFlow/PyTorch training through quantization and TensorFlow Lite deployment on Xilinx Zynq-7020 and Jetson Nano platforms. Key contributions include multi-sensor fusion via Python I2C/UART interfaces, computer vision-based optical navigation, and a four-tier safety architecture ensuring fault tolerance during communication blackouts. Validated through hardware-in-the-loop testing, this open-source stack enables sub-meter precision satellite rendezvous while maintaining real-time inference within severe power and memory constraints typical of nanosatellite missions.