Multimodal bipedal locomotion generation with passive dynamics via deep reinforcement learning

Area of Science:

• Robotics
• Control Systems
• Machine Learning

Background:

• Generating multimodal locomotion in underactuated bipedal robots is challenging due to diverse dynamical modes.
• Effective utilization of body morphology is crucial for energy-efficient locomotion in such systems.

Purpose of the Study:

• To develop a framework for reproducing multimodal bipedal locomotion using passive dynamics.
• To apply deep reinforcement learning (DRL) for controlling underactuated bipedal robots.

Main Methods:

• An underactuated bipedal model was developed, inspired by passive walkers.
• A DRL controller was designed and trained using a curriculum learning method for reward function tuning.
• The model learned to adjust gaits using a single command input.

Main Results:

• The DRL framework successfully enabled multimodal locomotion, including walking and running.
• The bipedal model demonstrated the ability to transition between different gaits.
• Effective use of passive dynamics was achieved through the DRL controller.

Conclusions:

• Deep reinforcement learning is a viable method for generating diverse gaits in bipedal robots.
• The proposed framework facilitates energy-efficient locomotion by leveraging passive dynamics.
• This approach offers a promising direction for advanced robotic locomotion control.