LSWM: A Long-Short History World Model for Bipedal Locomotion via Reinforcement Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Bipedal robot locomotion in unstructured terrain is limited by sensor noise, missing states, and poor future prediction.
• Conventional methods rely on long-term history for state reconstruction, neglecting short-term dynamics and risk anticipation.

Purpose of the Study:

• To propose and evaluate a Long-Short World Model (LSWM) for enhancing bipedal robot locomotion.
• To improve robustness against sensor noise and enhance state observability using short-term history.
• To increase adaptability to complex environments through future state prediction.

Main Methods:

• Developed a Long-Short World Model (LSWM) framework with two modules: State Reconstruction Module (SRM) and Future State Prediction Module (SPM).
• SRM utilizes long-term history to reconstruct privileged information within short-term history, improving robustness and observability.
• SPM predicts future privileged information to enhance adaptability and risk anticipation.

Main Results:

• LSWM demonstrated significant performance advantages in complex and dynamic environments through extensive simulations and real-world experiments.
• Achieved a 94% success rate in an indoor stair-climbing task, outperforming state-of-the-art methods by at least 34%.

Conclusions:

• The proposed LSWM framework effectively enhances bipedal robot locomotion capabilities in unstructured and dynamic environments.
• LSWM offers a robust solution for real-world robotic applications requiring advanced state estimation and predictive control.