DynaWM: Dynamics-Aware Distillation with World Model and Momentum Targets for Smooth Locomotion over Continuous Stairs

Bipedal-wheeled robots often struggle with continuous staircase traversal because existing teacher-student distillation frameworks fail to capture comprehensive terrain geometry and suffer from unstable training dynamics. The authors investigate how to improve terrain perception and motion smoothness for these robots by enhancing the representation learning process.

The authors propose DynaWM, a framework that introduces two key innovations to the standard Concurrent Teacher-Student (CTS) paradigm. First, they integrate a world model as a regularizer for the teacher encoder, which forces the model to predict future states and thereby ensures that the learned latent representations capture essential forward-dynamics and terrain geometry rather than just immediate reward-maximizing features. Second, they implement a momentum target encoder—inspired by BYOL—to provide stable distillation targets for the student. This prevents the dimensional collapse that typically occurs when a student tries to learn from a non-stationary, rapidly updating teacher encoder.

Through Principal Component Analysis (PCA) and quantitative metrics, the authors demonstrate that their encoder captures terrain height information in a more hierarchical and structured manner than baseline methods. Experimental results in both simulation and real-world hardware tests show that DynaWM enables bipedal-wheeled robots to traverse continuous stairs with significantly improved motion smoothness and terrain adaptability. The ablation studies confirm that both the world model and the momentum target encoder are critical for preventing the loss of geometric information and maintaining training stability.