MGDP: Mastering a Generalized Depth Perception Model for Quadruped Locomotion.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Deep Reinforcement Learning (DRL) controllers excel on challenging terrains but lack generality and transferability.
• Existing methods suffer from high computational costs and sensor noise sensitivity.

Purpose of the Study:

• To develop a generalized control framework for robust and efficient quadrupedal locomotion.
• To overcome limitations in terrain generality, platform transferability, computational overhead, and sensor noise.

Main Methods:

• Proposed the Mastering a Generalized Contrastive Depth Model (MGDP) framework.
• Utilized NVIDIA Warp for efficient parallel depth image computation.
• Extracted low-dimensional terrain features from multi-modal inputs (depth images, height maps).
• Integrated depth map denoising and terrain-adaptive reward functions.

Main Results:

• MGDP decouples perception from dynamics, reducing memory and computational costs.
• Enabled complex locomotion skills (climbing, jumping, crawling, squeezing) in a single training stage.
• Achieved superior cross-terrain generalization and efficient fine-tuning across different robot morphologies.

Conclusions:

• MGDP advances unified, efficient, and generalized frameworks for quadrupedal locomotion control.
• The dynamics-decoupled perception model facilitates rapid adaptation to diverse robot platforms.