Motion trajectory prediction of quadruped robots in complex terrain by improving the transformer temporal modeling...

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Quadruped robot trajectory prediction in complex terrain is challenged by foot-to-ground contact transients and multi-scale temporal dependencies.
• Traditional Transformer models face limitations in simultaneously capturing short-term impacts and long-range gait patterns under real-time constraints.

Purpose of the Study:

• To propose an improved Transformer temporal modeling algorithm for enhanced motion prediction in complex terrain for quadruped robots.
• To validate the potential of Transformers in accurately modeling contact transients and gait patterns.

Main Methods:

• Integration of multi-scale hybrid attention mechanisms, specifically local-sparse global hybrid attention, for capturing both short-term impacts and long-range dependencies.
• Adaptation of non-uniform sampling with relative temporal encoding to highlight key transient features.
• Parallel fusion of information from different time domains using a dynamic gated feedforward network and a multi-scale temporal encoder.

Main Results:

• Achieved a short-term Final Displacement Error (FDE) of 0.018 m, a significant reduction compared to traditional Transformers.
• Demonstrated high accuracy on gravel (0.095 m FDE) and sloping terrains (0.130 m FDE).
• Maintained a low inference latency of 12.45 ms, balancing performance and accuracy.

Conclusions:

• The proposed framework effectively balances real-time performance and high accuracy for quadruped robot trajectory prediction in complex environments.
• The method significantly enhances the stability and reliability of future motion predictions for quadruped robots.