Fine-grained dexterous manipulation presents significant challenges for robots due to the need for precise object handling, coordination of contact forces, and utilization of visual observations. This research aims to address these challenges by investigating the integration of vision and kinesthetic sensors, sim2real techniques, and generalization through embodiment. The objective is to develop end-to-end algorithms and models that enable robots to manipulate objects with exceptional precision and adaptability. The research will focus on learning from large-scale data, transferring knowledge from simulations to real-world scenarios, and efficiently generalizing through low-shot fine-tuning.