Advancing Adaptive Autonomy through Procedural Space Environments
- Paper ID
100162
- DOI
- author
- company
University of Luxembourg;
- country
Luxembourg
- year
2025
- abstract
The expansion of sustainable human presence beyond Earth requires robust autonomous systems capable of reliable operation across extreme and unpredictable conditions of space. However, the scarcity of real-world space robotics datasets combined with the prohibitive cost of hardware testing necessitates innovative approaches for system development and validation. Our open-source simulation framework addresses these constraints through procedural generation and domain randomization techniques that create virtually unlimited environmental variations that span multiple application domains of space, as opposed to traditional methods that rely on static scenarios. The framework leverages Blender to synthesize 3D assets at a configurable level of detail from parametric pipelines, which can be randomized to generate a wide range of extraterrestrial landscapes, structures, spacecraft, and mission-critical features. These assets integrate seamlessly into existing simulators through automated export while maintaining visual fidelity by baking procedural materials into optimized textures. Additionally, we introduce a dedicated simulation framework built on NVIDIA Omniverse that provides space-relevant environments for testing various autonomous platforms. In addition to rigid-body and soft-body dynamics, the framework also incorporates particle-based physics to model granular mechanics during regolith interactions. To further enhance realism and challenge autonomous systems, all environments systematically randomize both physical parameters and visual aspects by varying conditions such as gravity, illumination, material properties and appearance. Moreover, our scalable architecture supports the parallel execution of thousands of unique environment instances, which simultaneously reduces validation cycle times and diversifies data collection. The performance of autonomous space systems can be evaluated across several key tasks, ranging from exploratory traversal over challenging terrain to precision assembly, resource excavation, and tumbling debris capture. It supports platforms essential for future space infrastructure, including fixed and mobile manipulators, wheeled rovers, legged robots, aerial vehicles, orbital spacecraft, and humanoids. Benchmarking with this setup reveals a fundamental challenge in generalization capabilities. While state-of-the-art reinforcement learning agents reliably achieve success rates above 90\% in static environments, performance significantly degrades when confronted with procedural variation. This generalization gap underscores the value of our framework in exposing critical robustness issues that would remain hidden until hardware deployment. The introduced simulation framework provides an essential building block for sustainable space exploration by enabling affordable validation of autonomous systems. Although simulated environments cannot be fully validated against the limited available extraterrestrial data, the diversity provided through procedural generation ensures coverage of plausible operational conditions far beyond what could be anticipated through manual design, thus advancing standardized approaches towards robust autonomous systems.