Learning robust perceptive locomotion for quadrupedal robots in the wild

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Legged robots require robust perception for autonomous operation in hazardous environments.
• Current exteroceptive perception methods struggle with challenging conditions like snow, vegetation, and poor lighting.
• Reliance on proprioception alone limits locomotion speed due to the need for physical terrain interaction.

Purpose of the Study:

• To develop a robust and general solution for integrating exteroceptive and proprioceptive perception in legged robots.
• To enhance the speed and stability of legged robot locomotion through improved sensory integration.
• To overcome the limitations of current perception methods in diverse and difficult environments.

Main Methods:

• An attention-based recurrent encoder was developed to integrate proprioceptive and exteroceptive sensory inputs.
• The encoder was trained end-to-end, enabling it to learn seamless fusion of different perception modalities.
• The system was designed to avoid heuristic-based integration, promoting a more general solution.

Main Results:

• The integrated perception system resulted in a legged locomotion controller with high robustness and speed.
• The controller demonstrated successful navigation in various challenging natural and urban environments across different seasons.
• The robot completed an hour-long hike in the Alps within human-recommended timeframes.

Conclusions:

• Integrating exteroceptive and proprioceptive perception via an attention-based encoder significantly improves legged robot locomotion.
• The developed method offers a robust and general solution for autonomous navigation in complex and unpredictable terrains.
• This advancement paves the way for more capable robotic exploration and operation in remote and hazardous areas.