Perceiver IO Model for Visual Servoing of Reflective Parts Based on Polarization Camera Feedback

Source abstract

Reflective surfaces challenge industrial visual servoing because specular highlights can obscure true edges, distort pose cues and destabilize closed-loop robot correction. This paper introduces a Perceiver IO model for visual servoing of reflective parts based on polarization camera feedback. The method constructs polarization intensity difference, angle- sensitive response, reflection confidence and valid-edge masks, then combines these visual cues with servo-error vectors and robot-state tags. Perceiver IO maps heterogeneous image and state tokens into latent representations for pose correction, confidence estimation and review-priority output. A reflection-aware confidence branch is added to reduce unstable control updates when highlights, saturation or partial occlusion weaken geometric evidence. A synthetic dataset of 6,055 servo window assemblies is constructed under weak reflection, medium reflection, strong reflection, partial occlusion and mixed glare. The model achieves a Servo Success Rate of 94.8%, a mean final Pose Error of 0.62 mm and an average convergence time of 1.04 s. Compared with the compact convolutional servo baseline, final pose error under strong reflection is reduced by 28.6%, and the unstable update rate decreases by 16.4%. These results indicate that polarization feedback, when fused with robot-state information through latent attention, improves reflective-part servo reliability while retaining traceable pose correction, confidence and review information for safe robot assembly.