DynaWM enables bipedal-wheeled robots to climb continuous stairs smoothly
World model distillation and momentum targets solve staircase traversal for robots.
Bipedal-wheeled robots have struggled with long staircase traversal because existing teacher-student control frameworks lose dynamics-aware representations and incomplete terrain encoding. To address this, researchers Haidong Hou, Zhangguo Yu, Hengbo Qi, and Jianlin Zhang introduce DynaWM, a dynamics-aware representation learning framework accepted at IROS 2026. DynaWM incorporates a world model as a regularizer to enforce forward-dynamics awareness, allowing the encoder to capture comprehensive terrain geometry while enabling transparent hierarchical visualization. Additionally, a momentum target encoder provides consistent distillation targets, preventing dimensional collapse caused by non-stationary teacher updates.
The team evaluated DynaWM using PCA visualization and quantitative metrics, showing that the encoder hierarchically captures terrain geometry with higher encoding capability than prior methods. Experiments in both simulation and on real hardware demonstrated superior terrain adaptability and motion smoothness, enabling the robots to traverse diverse continuous stairs reliably. This advance moves bipedal-wheeled robots closer to practical deployment in environments like multi-level buildings, where staircase negotiation is essential.
- Introduces a world model regularizer to preserve forward-dynamics awareness and terrain geometry encoding.
- Uses a momentum target encoder to stabilize knowledge distillation and prevent dimensional collapse.
- Demonstrated smooth traversal of continuous stairs in both simulation and real-world hardware experiments.
Why It Matters
Solves a key locomotion challenge for bipedal-wheeled robots, enabling practical use in multi-level real-world environments.