Multiple Instance Learning (MIL) is a powerful framework for robotic perception and decision-making, where labels are associated with sets of instances rather than individual data points. We propose gradient-based Set-Input Trees, a novel tree-based architecture for MIL that addresses both classification and regression, with significant potential for application in robotics. Unlike conventional methods relying on fixed aggregation (e.g., min/max pooling), the proposed architecture integrates gradient-based trees with an attention mechanism: instances are processed independently, while leaf embeddings are pooled via learned attention weights. Furthermore, a decision tree ensemble is trained as an aggregation function to handle bag-level embeddings. This preserves interpretability, crucial for ensuring safety and trust in robotic systems, while simultaneously capturing bag-level structure inherent in complex robotic environments. For regression tasks, common in robot control, we introduce a synthetic MIL formulation, feature-to-bag conversion, enabling evaluation on continuous targets. Experiments show outperformance of the proposed algorithms on standard MIL benchmarks comparing to tree-based models. The model’s tree-based design ensures scalability and transparency, bridging instance-level decisions with set-valued predictions. Codes implementing the proposed algorithms are publicly available.