Proprioceptive-based terrain-adaptive rolling control for wheeled-quadruped robots

This paper presents a proprioception-based rolling control framework for wheeled-quadruped robots. The proposed framework serves as a reactive control approach for adaptive terrain traversal and body stabilization, without explicit motion planning. For wheeled-quadruped robots, adding wheels significantly increases the inertia of the legs. As a result, the accuracy of the Single Rigid Body Dynamics (SRBD) model, a method commonly used for quadruped robots, is compromised. Meanwhile, multibody dynamic models are computationally expensive and often require high-performance onboard hardware. In addition, most current wheeled-quadruped robots rely on stepping-based steering, which increases energy consumption and induces large body oscillations. To address these limitations, we introduce a Two-Layer Lumped-Mass (TLLM) model and develop a hierarchical controller that combines Quadratic Programing (QP) with Virtual Model Control (VMC). By reducing the problem dimension and linearizing the dynamics, the controller only needs to solve a single linear optimization problem, without imposing any additional wheel-rolling constraints. This reduces the computational cost while retaining adequate tracking performance of body attitude and wheel-end positions, thereby supporting active-suspension control and differential steering. The proposed method is validated on the Lite3-W robot through both simulation and hardware experiments. During traversal of a unilateral obstacle, stairs, a slope, randomized obstacles, staggered wavy obstacles, and outdoor rough terrain, the RMSE of body attitude angles remains below 0.045 rad. Meanwhile, the average RMSE of the four wheel position errors in both the x -and y -directions remains below 23.854 mm. For differential steering, the Cost of Transport (CoT) is reduced by 44% compared with stepping-based steering. In contrast, the SRBD-based baseline fails to complete the corresponding tests due to poor wheel-end position tracking. In addition, the simulation-based comparison with a multibody dynamics baseline shows that the proposed approach improves computational efficiency by more than a factor of four.