Latent Space Representations for Marker-Less Realtime Hand-Eye Calibration.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Marker-less hand-eye calibration is crucial for robot-unstructured environment interaction.
• Monocular cameras offer cost-effectiveness but struggle with coordinate projection accuracy.
• Existing learning-based methods often fail in orientation prediction for spatial transforms.

Purpose of the Study:

• To introduce a novel hand-eye calibration procedure for robots using marker-less techniques.
• To overcome orientation domain limitations in learning-based spatial transform regression.
• To enable accurate robot-sensor transformation in diverse and challenging conditions.

Main Methods:

• Utilized an augmented autoencoder neural network to infer rotation representations.
• Analyzed latent space vectors from the autoencoding process to improve orientation accuracy.
• Employed a color-depth camera for data acquisition and point cloud registration for evaluation.

Main Results:

• The proposed method achieves computationally inexpensive and real-time hand-eye calibration.
• Demonstrated robustness in varied lighting and occlusion conditions.
• Achieved comparable or improved accuracy against traditional marker-based methods.

Conclusions:

• The autoencoder-based approach effectively addresses orientation challenges in marker-less hand-eye calibration.
• This technique offers a viable, efficient solution for robot-sensor calibration in unstructured settings.
• The method shows promise for real-time applications requiring precise robotic manipulation.