Path Planning Generator with Metadata through a Domain Change by GAN between Physical and Virtual Environments.

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Robotic systems require advanced perception for real-time interaction, often necessitating expensive specialized equipment.
• Path planning in intelligent robots involves complex subsystems (perception, localization, control) demanding rapid responses, which increases development costs and time.
• Virtual reality (VR) is used for algorithm training, generating synthetic data that can be integrated with real-world data.

Purpose of the Study:

• To investigate cost-reduction strategies for robotic perception systems.
• To explore the use of Generative Adversarial Networks (GANs) for integrating virtual and real-world data.
• To evaluate a metadata-driven approach for describing agent properties in robotic environments.

Main Methods:

• Utilized Generative Adversarial Networks (GANs) to connect virtual datasets with real-world data.
• Implemented a metadata approach to describe agent properties within an environment.
• Tested the system with an augmented reality (AR) system and a micro aerial vehicle (MAV) on embedded devices.

Main Results:

• The GAN-based approach reduced development time and reliance on extensive physical world samples.
• The metadata approach, tested on AR and MAV systems, demonstrated feasibility in real environments.
• The system's performance was evaluated using metadata, showing potential for resource and cost reduction.

Conclusions:

• Generative Adversarial Networks (GANs) offer a viable method to reduce costs and development time in robotic perception by leveraging virtual data.
• Metadata-driven descriptions are effective for evaluating robotic system performance in real-world scenarios.
• External factors like illumination variations pose limitations for camera-dependent systems, highlighting areas for future improvement.