Navigating to objects in the real world.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Classical spatial navigation pipelines lack semantic understanding for real-world robot deployment.
• Learning-based approaches, including end-to-end and modular methods, aim to enhance robot navigation.
• Previous evaluations of visual navigation policies have been limited, primarily in simulation.

Purpose of the Study:

• To empirically compare semantic visual navigation methods in real-world uncontrolled environments.
• To evaluate the performance of classical, modular, and end-to-end learning approaches on mobile robots.
• To identify challenges in current simulation benchmarks for robot navigation.

Main Methods:

• A large-scale empirical study comparing semantic visual navigation methods.
• Methods tested include classical, modular learning, and end-to-end learning approaches.
• Evaluation conducted across six homes without prior mapping or instrumentation.

Main Results:

• Modular learning achieved a 90% success rate in real-world semantic navigation.
• End-to-end learning demonstrated a significant performance drop from 77% in simulation to 23% in the real world.
• The primary cause for end-to-end failure was identified as a large image domain gap between simulation and reality.

Conclusions:

• Modular learning is a reliable approach for real-world robot navigation to objects, enabling effective sim-to-real transfer.
• Current simulators are unreliable benchmarks due to significant sim-to-real gaps in images and error modes.
• Recommendations are provided for improving simulators and advancing research in semantic visual navigation.