Bio-inspired sparse expert control for agile and adaptive wheeled-legged locomotion

Wheeled-legged robots offer agile mobility, yet controlling them robustly remains challenging due to the intrinsic conflict between continuous rolling and discrete stepping dynamics. Conventional controllers often struggle to reconcile these hybrid objectives, leading to instability on heterogeneous terrains. To address this, we propose a bio-inspired, residual-stabilized sparse expert framework for wheeled-legged hybrid locomotion. Unlike standard MoE approaches, the proposed architecture incorporates a cerebellum-inspired Residual Reflex Module and inhibition-guided regularization to stabilize expert transitions and suppress disturbances during rolling, stepping, and contact switching. By decoupling strategy selection from low-level stabilization, the controller autonomously adapts locomotion behavior and transitions between wheel-dominant rolling and leg-assisted stepping using only directional velocity commands. Extensive simulation and real-world testing on the DEEPRobotics Lynx platform demonstrate strong zero-shot transfer capability. Payload-carrying tests are further conducted to evaluate deployment robustness under representative operational conditions. The controller successfully traverses extreme terrains, including 50° slopes and 25 cm steps, achieving over 90% consistency with simulation limits. The proposed method also reduces the Cost of Transport (CoT) by 29.7% compared with the adapted PPO baseline, indicating improved locomotion efficiency for defence and rescue-oriented mobility tasks.