My research develops autonomous guidance solutions for low-thrust interplanetary CubeSat missions. CubeSats have democratized access to deep space thanks to low cost and rapid development cycles, but the operational infrastructure has not scaled with this growth. In particular, the Deep Space Network is increasingly saturated, motivating a shift from ground-centric operations to onboard autonomy. At the same time, CubeSats are constrained by mass, power and thermal limits that restrict onboard processing capability, so any autonomy must be both robust and computationally efficient.

The work addresses these challenges by designing a closed-loop guidance architecture able to plan and execute low-thrust transfers in full autonomy. The core approach integrates Deep Reinforcement Learning (DRL) with Sequential Convex Programming (SCP) inside a high-reliability nonlinear Model Predictive Control framework. A policy-based DRL agent is trained to reproduce fuel-optimal, low-thrust trajectories while remaining adaptive to disturbances and the stochastic occurrence of actuator faults. The trained policy provides near-optimal control actions in real time on constrained hardware.

To secure the terminal accuracy required for planetary capture or precise flybys, the DRL output is used as a high-quality initial guess for a Sequential Convex Optimization stage. This hybrid DRL–SCP synergy reduces the sensitivity of classical solvers to initial conditions, improving convergence speed and success rate while preserving the optimality properties of convex methods. The optimal, discretized thrust profile is then regularized to yield a continuous control law suitable for flight, and the entire pipeline runs in a receding-horizon fashion to enable closed-loop corrections throughout the transfer.

Robustness and performance are demonstrated through extensive Monte Carlo campaigns covering challenging scenarios such as transfers between Earth and Mars, and rendezvous with Near-Earth Objects. To bridge toward flight readiness, the framework is being validated with Processor-in-the-Loop simulations that measure real-time execution on representative embedded processors, quantifying computational overhead, memory usage and convergence reliability. Planned Hardware-in-the-Loop tests will further assess interactions with physical sensors and actuators in a simulated deep-space environment.

Ultimately, this research aims to provide a scalable path toward deep-space CubeSat missions that can navigate the solar system with minimal ground intervention. By combining learning-based adaptability with optimization-based rigor and by validating performance on embedded platforms, the project targets autonomy that is both mission-effective and flight-ready.

Click here to view my poster.