Work

Decoupled spatio-temporal modeling for high-fidelity lunar robot sensorimotor forecasting

Year: 2026

Type: article

Abstract: • Adaptive semantic planning via multi-granularity knowledge graph construction and ontology-driven reasoning, enabling contextual understanding and task decomposition in complex lunar south pole environments. • Closed-loop ”Plan–Execute–Reflect–Iterate” mechanism for dynamic decision optimization, allowing robust task adjustment under uncertain and evolving physical and operational constraints. • High-fidelity lunar surface simulation environment incorporating PSR-specific conditions to validate planning accuracy, robustness, and efficiency in realistic mission scenarios. Accurate modeling of robotic sensorimotor sequences is critical for autonomous systems, yet it presents a dual challenge: capturing complex inter-variable couplings and long-range temporal dependencies. The inherent entanglement of these distinct dynamics makes it difficult for a single, unified model to address both effectively. To tackle this core issue, we propose a decoupled modeling paradigm that handles these challenges with specialized modules. As an architectural embodiment of this principle, we introduce StambaNet, a deep learning architecture that features a parallel dual-branch design. An improved STAR module, enhanced with multiscale pyramid fusion, is dedicated to modeling instantaneous cross-variable correlations, while a Mamba module efficiently captures long-range dependencies. These specialized features are dynamically integrated via an adaptive gating mechanism and trained end-to-end with a self-supervised masked modeling task. On a custom dataset of four challenging lunar robotic tasks, our approach significantly outperforms five baseline models, with the largest gains in tasks with high inter-variable complexity. These results validate the effectiveness of our decoupled design. Furthermore, ablation studies, including a branch-sensitivity analysis, empirically validate that the two branches possess complementary inductive biases, which explains the model’s superior accuracy and robustness.

Source: Advances in Space Research

DOI: 10.1016/j.asr.2026.01.026

Authors: Ziliang Zhao, Yiling Kuang, Cheng Wei, Xibin Cao

Institution: Harbin Institute of Technology

Language: English

Topic: Reinforcement Learning in Robotics

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $2,780

Funders: National University's Basic Research Foundation of China, Fundamental Research Funds for the Central Universities