Smooth robot control without oscillations or extra complexity

Robots trained with reinforcement learning often produce shaky, high-frequency control signals that cause instability and safety problems in the real world. Existing fixes require predicting long action sequences, which bloats the network and breaks standard training. This paper introduces Dual-Window Smoothing (DWS), which enforces smooth control by constraining how actions can change between steps rather than expanding the output size. Tests on robot control benchmarks and autonomous driving show DWS produces steadier, safer behavior than competing methods.