External randomness exists in all sports but is perhaps most obvious in golf putting where robotic putters sink only 80% of 5 m putts due to unpredictable ball-green dynamics. The purpose of this study was to test whether physical randomness training can improve putting performance in novices. A virtual random-physics golf-putting game was developed based on controlled ball-roll data. Thirty-two subjects were assigned a unique randomness gain (RG) ranging from 0.1 to 2.0-times real-world randomness. Putter face kinematics were measured in 5 m laboratory putts before and after five days of virtual training. Performance was quantified using putt success rate and "miss-adjustment correlation" (MAC), the correlation between left-right miss magnitude and subsequent right-left kinematic adjustments. Results showed no RG-success correlation (r = -0.066, p = 0.719) but mildly stronger correlations with MAC for face angle (r = -0.168, p = 0.358) and clubhead path (r = -0.302, p = 0.093). The strongest RG-MAC correlation was observed during virtual training (r = -0.692, p < 0.001). These results suggest that subjects quickly adapt to physical randomness in virtual training, and also that this learning may weakly transfer to real golf putting kinematics. Adaptation to external physical randomness during virtual training may therefore help golfers adapt to external randomness in real-world environments.