Autonomous CubeSat Docking

This research presents a deep reinforcement learning framework for autonomous spacecraft rendezvous and docking using Proximal Policy Optimization (PPO) with curriculum learning. The system trains an RL agent in a custom 3D orbital mechanics environment incorporating gravitational perturbations, atmospheric drag, and fuel constraints. Key Methodology: Algorithm: Proximal Policy Optimization (PPO) with clipped surrogate objective Training: Three-phase curriculum learning (Easy → Normal → Hard) Reward Shaping: Multi-objective optimization for fuel efficiency, velocity alignment, and soft-docking Environment: 8-dimensional observation space with 3D thrust control actions Validated Performance (50 episodes): Success Rate: 100% (50/50) Mean Convergence Time: 25.8 steps Fuel Efficiency: 6.97 units consumed (93.03% remaining) Position Precision: 0.905m ± 0.05m final distance Approach Velocity: 0.223 m/s (conservative, safe) The system includes a VPython-based pilot interface providing real-time 3D visualization with heads-up display (HUD) elements for human-in-the-loop monitoring. The trained policy achieves exceptional consistency (σ=0.61 fuel units) and demonstrates emergent behaviors including efficient coasting arcs and conservative terminal approach velocities well below safety thresholds.