Orbit-Planner: Towards Latent World Models for On-Orbit Obstacle Avoidance of Satellite Agents
| Source: arXiv AI
Tags: satellite AI, space robotics, world models, obstacle avoidance, reinforcement learning
Orbit-Planner applies latent world models to satellite on-orbit navigation, achieving 91.7% obstacle avoidance success in Isaac Sim by learning action-conditioned spacecraft dynamics without predefined maps.
Details
Researchers from the Chinese Academy of Sciences present Orbit-Planner, a two-stage latent world model for satellite obstacle avoidance in on-orbit scenarios. The system addresses a gap in conventional space planners: relying on predefined maps and fixed environmental assumptions makes them brittle when environments are dynamic or poorly mapped. Orbit-Planner learns action-conditioned spacecraft dynamics and performs future-state rollouts in latent space — predicting what will happen if a given maneuver is executed — without needing an explicit geometric map of the environment. A Physics Probe module decodes physical state changes (position, velocity) from imagined latent trajectories, bridging the abstraction gap between compressed representations and interpretable physics. In closed-loop obstacle-avoidance navigation tests in NVIDIA Isaac Sim, the system achieves 91.7% success rate. The paper (4 pages) was accepted to AP-GARSS 2026, a geoscience and remote sensing conference. This is a niche but genuine contribution at the intersection of AI and space systems. The short paper format limits methodological depth, and Isaac Sim results don't yet demonstrate hardware validation.