Safe DRL with CBF filter guarantees spacecraft reorientation safety
New DRL method uses a control barrier function to ensure safe spacecraft maneuvers...
A new paper on arXiv (2605.19967) from researchers Juntang Yang and Mohamed Khalil Ben-Larbi presents a safe deep reinforcement learning (DRL) framework for spacecraft reorientation control. The challenge is to reorient a spacecraft while keeping its sensitive instruments away from bright celestial objects (a pointing keep-out zone). The authors introduce a novel state space representation that compactly encodes the attitude constraint zone, paired with a reward function designed to balance the control objective and constraint enforcement.
The approach uses the Soft Actor-Critic (SAC) algorithm to handle continuous state and action spaces, and a curriculum learning strategy to progressively train the agent. Crucially, a control barrier function (CBF)-based safety filter is added during deployment to guarantee constraint satisfaction—Monte Carlo simulations show that reward shaping alone cannot reliably ensure safety, while the CBF filter provides formal guarantees. This work demonstrates a practical pipeline for integrating learning-based control with formal safety methods in aerospace applications.
- Uses Soft Actor-Critic (SAC) for continuous control of spacecraft attitude
- Curriculum learning speeds up training for complex reorientation tasks
- CBF safety filter ensures 100% constraint compliance vs reward shaping alone
Why It Matters
Enables AI-driven spacecraft maneuvering with provable safety guarantees, critical for autonomous satellite operations and deep space missions.