Meta-RL Enhances Spacecraft Safety and Fuel Efficiency in Adversarial Scenarios

Autonomous spacecraft rendezvous and proximity operations (RPO) demand controllers that can guarantee safety while adhering to thrust constraints and minimizing fuel consumption. Input-constrained control barrier functions (ICCBFs) offer a method for ensuring safety in nonlinear systems with actuation limits. Previous work demonstrated that meta-reinforcement learning (meta-RL) could robustly learn ICCBF class-K functions for RPO. This research extends that framework by comparing the performance of three recurrent network architectures—Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU), and Selective State Space Model (Mamba)—and two training algorithms, Proximal Policy Optimization (PPO) and Soft Actor Critic (SAC). The goal was to identify the optimal configuration for tuning ICCBF class-K functions via meta-RL. Performance was assessed in both cooperative and adversarial scenarios, where a target spacecraft might act to compromise the chaser spacecraft's safety. The results indicate that state space models like Mamba, when combined with PPO, deliver superior task completion, safety guarantees, and fuel efficiency across all tested cooperative and uncooperative situations.