RoboTAG: End-to-end Robot Configuration Estimation via Topological Alignment Graph

The introduction of RoboTAG marks a significant advancement in robot pose estimation, a critical challenge in robotics and computer vision. Traditional methods often rely on labeled data, which can be limited in real-world applications, leading to a sim-to-real gap. RoboTAG innovatively combines a 3D branch with a 2D branch, allowing for the co-evolution of these representations and reducing the dependency on labeled training data. This method utilizes in-the-wild images without annotations, making it more adaptable and practical. Experimental results indicate its effectiveness across different robot types, suggesting a promising future for its application in real-world scenarios.