Task-Oriented Humanoid Autonomous Surveying via Uncertainty-Aware Neural Field Reconstruction and Whole-Body Model Predictive Control

Humanoid robots operating in industrial environments require autonomous surveying capabilities that jointly support scene understanding, safe locomotion, and manipulation-oriented perception. However, existing active reconstruction and model predictive control methods are typically separated, making it difficult to propagate perception uncertainty into motion decisions or prioritize observations according to downstream manipulation value. This paper proposes UNF-WMPC, an uncertainty-aware neural-field whole-body model predictive control framework integrating semantic neural-field reconstruction, task-weighted information gain, receding-horizon viewpoint optimization, and whole-body safety constraints. The method models reconstruction uncertainty, predicts its reduction under candidate future observations, and incorporates uncertainty-dependent clearance, balance, joint, collision, and human-separation constraints into the control loop. Experiments on four indoor datasets show that UNF-WMPC consistently improves reconstruction quality, information-acquisition efficiency, surveying safety, and robustness under progressive perception degradation compared with active reconstruction baselines. These results demonstrate that coupling task-oriented perception with predictive whole-body control provides an effective foundation for reliable humanoid surveying and subsequent industrial manipulation.