Real-Time Terrain Recognition for Quadruped Robots Using Proprioceptive Sensors and Temporal Convolutional Networks

Area of Science:

• Robotics
• Machine Learning
• Sensor Fusion

Background:

• Quadruped robots require accurate terrain recognition for safe navigation in complex environments.
• External sensors face limitations due to lighting, occlusion, and surface reflectivity.
• Proprioceptive sensors offer a robust alternative for perception.

Purpose of the Study:

• To develop a real-time terrain recognition and slip estimation method for quadruped robots.
• To reduce reliance on external sensors by utilizing proprioceptive data.
• To improve the safety and adaptability of quadruped robot locomotion.

Main Methods:

• Utilized proprioceptive sensors for data acquisition.
• Developed a complementary perception module incorporating Temporal Convolutional Networks (TCNs).
• Validated the framework through extensive real-world experiments and embedded edge computing deployment.

Main Results:

• Achieved 98.8% terrain recognition accuracy with the proposed TCN method.
• Demonstrated superior performance compared to baseline models.
• Analyzed the impact of locomotion speed and environmental conditions on slip detection.

Conclusions:

• The TCN-based proprioceptive sensing method provides reliable real-time terrain recognition and slip estimation.
• The system enables quadruped robots to detect terrain types and surface states for adaptive locomotion.
• The proposed solution is cost-effective, robust, and low-latency, suitable for diverse terrains.