This　paper　presents　an　interactive　motion　control　method　based　on　reinforcement　learning,　designed　to　assist　children　with　autism　who　have　social　motor　impairments　through　a　mirror　game　intervention.　The　virtual　teacher　uses　the　Deep　Deterministic　Policy　Gradient　(DDPG)　algorithm　to　optimize　its　actions,　guiding　the　participant　to　follow　a　Lissajous　trajectory.　To　ensure　safety,　a　motion-correction　mechanism　was　developed,　which　automatically　adjusts　actions　when　the　predicted　trajectory　surpasses　predefined　safety　boundaries.　The　reward　function　considers　both　the　distance　between　the　virtual　teacher　and　the　target　trajectory,　as　well　as　the　distance　between　the　virtual　teacher　and　the　participant,　with　dynamic　adjustments　applied　by　the　motion-correction　mechanism.　Experimental　results　demonstrate　that　the　virtual　teacher　effectively　guides　the　participant　towards　the　target　trajectory　while　adhering　to　safety　constraints.　©　2025　IEEE.