Attention-based map encoding for learning generalized legged locomotion

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Dynamic locomotion in legged robots is crucial for expanding mobile robot capabilities but faces challenges in precision and robustness across diverse terrains.
• Traditional model-based controllers lack robustness to real-world uncertainties, while learning-based controllers may lack precision on sparse terrains.
• Hybrid methods combine approaches but are computationally intensive and limited by model-based planners.

Purpose of the Study:

• To develop a generalized legged locomotion controller that is robust to uncertainties and precise on diverse, sparse terrains.
• To leverage attention mechanisms and reinforcement learning for improved topographical perception and foothold planning.
• To enable agile and robust dynamic locomotion for legged robots in complex environments.

Main Methods:

• Proposed an attention-based map encoding conditioned on robot proprioception, trained using reinforcement learning.
• Developed a novel controller integrating neural network-based topographical perception with dynamic locomotion planning.
• Trained controllers for both quadrupedal and humanoid robots, demonstrating adaptability.

Main Results:

• The attention-based network learned to effectively identify steppable areas for future footholds during dynamic navigation.
• Synthesized behaviors demonstrated robustness against uncertainties and enabled precise, agile traversal of sparse terrains.
• The method provided interpretable insights into the neural network's topographical perception.

Conclusions:

• The proposed attention-based controller achieves generalized and robust dynamic locomotion for legged robots across diverse and challenging terrains.
• This approach enhances precision and agility, overcoming limitations of traditional and purely learning-based methods.
• Real-world testing on quadrupedal and humanoid robots validated the controller's effectiveness in various scenarios, including unseen conditions.