EchoPT: A Pretrained Transformer Architecture That Predicts 2D In-Air Sonar Images for Mobile Robotics

Area of Science:

• Robotics
• Artificial Intelligence
• Sensor Fusion

Background:

• Perception is modeled as minimizing prediction error between internal world models and sensory input.
• Implementing predictive processing with sparse sonar data presents significant challenges for robotics.

Purpose of the Study:

• To develop a robust world model for robots using only ultrasound data.
• To enable predictive perception for enhanced robot behavior in unstructured environments.

Main Methods:

• Introduced EchoPT, a pretrained transformer architecture for predicting 2D sonar images.
• Utilized previous sensory data and robot ego-motion for predictions.
• Compared EchoPT performance against state-of-the-art techniques.

Main Results:

• EchoPT effectively predicts sonar images, overcoming challenges of sparse sonar reflection models.
• Demonstrated the model's capability in enabling predictive perception for robotic tasks.
• Achieved competitive performance compared to existing methods.

Conclusions:

• Predictive perception using transformer models like EchoPT is viable for robot navigation with ultrasound.
• This approach enhances robot autonomy by relying solely on sonar exteroception.
• EchoPT facilitates robust robot behavior in scenarios where other sensors are unavailable.