A Novel Obstacle Traversal Method for Multiple Robotic Fish Based on Cross-Modal Variational Autoencoders and...

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Precision control of multiple robotic fish in complex underwater environments presents significant challenges.
• Existing methods struggle with stable visual navigation and obstacle traversal for multi-robotic systems.

Purpose of the Study:

• To propose and validate a novel multi-robotic fish obstacle traversal technique.
• To enhance the visual navigation capabilities of robotic fish in complex underwater settings.

Main Methods:

• Utilized a cross-modal variational autoencoder (CM-VAE) to encode first-person robotic fish views into a low-dimensional latent space.
• Employed imitation learning to map latent features to robotic fish velocity commands.
• Conducted experiments with single and multiple robotic fish on linear, S-shaped, and circular gate frame trajectories.

Main Results:

• The proposed visual navigation method demonstrated stable traversal of various gate frame trajectories.
• The CM-VAE and imitation learning approach showed superior performance compared to end-to-end learning and unsupervised image reconstruction.
• Successfully enabled intelligent navigation for robotic fish in complex environments.

Conclusions:

• The developed technique offers a robust solution for multi-robotic fish visual navigation.
• This approach provides a new pathway for intelligent underwater robotic systems.
• The method significantly improves obstacle traversal and navigation stability in challenging underwater conditions.