This paper introduces a novel framework for robotic pruning systems that leverages deep evidential learning to provide calibrated uncertainty estimates for 3D point cloud segmentation and collision detection, addressing model, data, and sensor uncertainties.
Key findings
Introduces a unified uncertainty quantification framework integrating epistemic, aleatoric, and temporal uncertainties.
Presents an evidential deep learning architecture extending PointNet++ for uncertainty-aware 3D point cloud segmentation.
Develops a collision probability estimation module to transform segmentation uncertainties into actionable safety margins.
Includes comprehensive experimental validation with synthetic datasets, real-world evaluations, and ablation studies.
Limitations & open questions
Further research is needed to scale the framework for different agricultural environments and tree species.