Adaptive multi-mode locomotion for bipedal wheel-legged robots via sparse mixture-of-experts deep reinforcement...

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Bipedal wheel-legged robots offer combined benefits of wheeled efficiency and legged adaptability.
• Seamlessly transitioning between wheeled and legged locomotion in a unified control framework is a significant challenge.

Purpose of the Study:

• To develop a reinforcement learning control framework for smooth transitions between wheeled and legged locomotion in bipedal robots.
• To enhance training stability and performance using a Mixture of Experts (MoE) architecture.

Main Methods:

• Implemented a reinforcement learning control framework integrating the Mixture of Experts (MoE) architecture.
• Utilized a dynamic gating network and Top-K sparse activation for automatic motion mode allocation.
• Compared the MoE-enhanced algorithm against a single-network Proximal Policy Optimization (PPO) method.

Main Results:

• The MoE-enhanced algorithm demonstrated superior training stability and higher rewards compared to single-network PPO.
• The learned policy achieved smooth rolling on flat surfaces and dynamic leg-lifting gaits over obstacles.
• Significantly higher success rates were observed across various test terrains compared to single-network PPO.

Conclusions:

• The proposed MoE reinforcement learning framework effectively addresses the challenge of heterogeneous motion mode transitions in bipedal wheel-legged robots.
• This approach enhances robot adaptability and performance in complex and varied environments.
• The MoE architecture provides a robust solution for unified control of dual-mode locomotion.