Reinforcement Learning Framework Enables Adaptive Multi-Task Control for Bipedal Soccer Robots

A modular reinforcement learning framework has been developed to address motion stability and control switching challenges in bipedal soccer robots operating in dynamic combat environments. The system combines an open-loop feedforward oscillator with reinforcement learning-based feedback residuals, effectively separating basic gait generation from complex football actions. A posture-driven state machine clearly switches between the ball seeking and kicking network (BSKN) and the fall recovery network (FRN), preventing state interference. The FRN is trained through a progressive force attenuation curriculum learning strategy. This architecture was verified in Unity simulations, tackling issues related to deep coupling of multiple tasks and transitions between states like upright walking and fall recovery. The approach fundamentally addresses control switching problems between different operational states.