Abstract Visible light positioning (VLP) is a promising technique to bring location-based service for numerous Internet of Things applications. Recent advances in VLP have shown that machine learning (ML)-based positioning algorithms show satisfying performance in physical environments under highly noisy and interference-rich conditions. With so many ML methods proposed, one major concern is that trained models could fail due to environmental heterogeneity. In this paper, we propose AdVLP, a novel adversarial training method for VLP based on deep neural networks, to address the issue of the vulnerability of data-driven approaches, which happens when channel parameters change. Our proposed method, which is inspired by generative adversarial networks, manages to adapt the two domains of the source dataset and the target dataset. The channel parameters, e.g. the Lambertian order of light-emitting diode (LED) lights and a photodiode, the LED power as well as the location error of the lights, were discussed. We observed in experiments that the accuracy of a neural network method would decrease as the bias of the parameters became larger. Secondly, a higher dimension of the parameter change would make the neural network method more vulnerable. However, our proposed approach could achieve substantial improvement in positioning accuracy.