Aiming　at　solving　the　dynamic　incertitude　of　the　robot　soccer　game,　a　learning　method　based　on　Locally　Weighted　Regression　is　proposed　to　forecast　the　motion　of　the　ball,　especially　after　the　ball　bump　into　wall.　The　error　is　adjusted　with　fuzzy,　that　is,　the　robot　moves　based　on　the　forecasted　condition　of　the　ball.　Its　model　is　simple　for　the　decision-making　system　of　robot.　Its　usability　is　proved　by　the　simulation.