Autonomous free-flying robots require versatile control policies that can adapt to varying mission objectives under strict resource constraints. We introduce HYPER-GNC, a framework that utilizes hypernetworks to map semantic embeddings to the weight space of a deep actor-critic policy. This approach enables a single, compact controller to master a diverse suite of Guidance, Navigation, and Control tasks (track velocities, docking, navigation with obstacles, and inspection) without the need for task-specific retraining. We demonstrate that our semantic manifold formulation facilitates zero-shot generalization to novel mission profiles. Extensive experimental results show that HYPER-GNC maintains stability under significant physical perturbations and external body wrenches. Finally, we demonstrate the bridge from simulation to reality by deploying our policy on a physical satellite emulator. To support reproducibility, our code and trained models are made publicly available.
git clone
cd Hyper-GNC
./docker/container.py build
./docker/container.py start
./scripts/reinforcement_learning/rsl_rl/control_train_hypernet.sh
You can download the weights form the google drive and add them inside the logs folder of the docker. Download link.
./scripts/reinforcement_learning/rsl_rl/control_eval_hypernet.sh