Occluded object detection and exposure in cluttered environments with automated hyperspectral anomaly detection

Area of Science:

• Robotics
• Computer Vision
• Spectroscopy

Background:

• Cluttered environments with occlusions challenge robot manipulation.
• Identifying anomalous objects is difficult using spatial features alone.
• Hyperspectral imaging offers material-specific data for improved object recognition.

Purpose of the Study:

• To develop an automated hyperspectral anomaly detection method for cluttered robot workspaces.
• To improve robot perception and manipulation in complex environments.
• To detect fractional object presence without extensive labeled data.

Main Methods:

• Created a multi-modal data representation pairing depth and hyperspectral data.
• Developed an unsupervised autoencoder for anomaly detection based on reconstruction error.
• Applied motion primitives to expose anomalies for enhanced robot interaction.

Main Results:

• Successfully demonstrated anomaly detection in diverse cluttered environments.
• Consistently increased the exposed surface area of anomalies across four scenarios.
• Validated the utility of multi-modal anomaly detection for robot manipulation.

Conclusions:

• Hyperspectral sensing combined with depth data significantly enhances robot perception in cluttered scenes.
• The proposed method automates anomaly detection, improving the ability to identify partially obscured objects.
• This approach advances robot manipulation by enabling more robust object identification and interaction.