Robust Visual Embodiment

Self-supervised robotic self-modeling enables machines to autonomously infer their morphology and kinematics directly from visual data. However, existing approaches often fail in realistic conditions with visual noise and cluttered backgrounds. This paper addresses these limitations by introducing a robust visual embodiment framework.

The proposed pipeline integrates a task-aware denoising framework with semantic segmentation. 1. Semantic Segmentation: Uses a deep convolutional network to isolate the robot from complex backgrounds, overcoming the limitations of color-based segmentation. 2. Task-Aware Denoising: Combines Wiener filtering (for blur), Median filtering (for salt-and-pepper noise), and Non-Local Means with Intuitionistic Fuzzy Twin SVM (for Gaussian noise) to restore image quality while preserving morphological features. 3. Self-Modeling Engine: Feeds the processed visual data into the FFKSM framework to reconstruct the robot's kinematic model.

Extensive experiments on both simulated and physical robots demonstrated that the proposed framework significantly improves robustness. - Segmentation: Achieved over 4x improvement in IoU and F1-score compared to baseline methods in cluttered environments. - Morphology Reconstruction: Restored near-baseline accuracy in morphology prediction even under severe noise conditions where standard pipelines failed. - Real-World Validation: Validated on a custom 3D-printed 4-DOF manipulator, proving the system's effectiveness in real-world scenarios.