MOP: A Multimodal Object-Aware Policy for Robotic Manipulation via Geometry-Guided Fusion and Trajectory Prediction
3D imitation learning has demonstrated capability in diverse visuomotor tasks but often struggles with small objects or high-precision manipulation due to geometric sparsity. To address this, we propose the Multimodal Object-aware Policy (MOP), a novel framework that incorporates a geometry-guided fusion module to adaptively integrate 2D semantic features with 3D geometry for precise control. Additionally, we introduce a lightweight Object Position Prediction (OPP) module that serves as an auxiliary supervision signal. The training labels for this module are generated by using a cost-effective vision-based software tracker, effectively replacing expensive hardware-based motion capture systems. We evaluate our policy across 56 tasks on 3 simulation benchmarks. Experimental results demonstrate that MOP significantly outperforms baselines, achieving higher success rates with lower variance and efficient inference. Real-world experiments on 4 tasks further validate the robustness and transferability of our approach.