<p>Comprehending dynamic behavior of humans is vital, if we want to predict collision-free trajectories for autonomous moving platforms (such as driverless vehicles and social robots) in an environment where they coexist with humans. The task of generating collision-free trajectories is difficult, since the trajectories of human are inherently multimodal and their movements are bound to change based on interactions with neighborhood habitat. Hence, in order to output accurate trajectories, it is important to model efficient social spatio-temporal interactions among the agents. Recently, a considerable body of scientific inquiry has been conducted in this field, incorporating various attention mechanisms in modeling interactions among agents through generative learning. However, all the existing works are not effectively modeling the interactions among agents, leading to collisions in complex scenarios. In our research work, we tried to buckle down this problem by designing an architecture titled “dynamic modeling of collision-free trajectories using generative adversarial networks” (DMCT-GAN), which can efficiently model social spatio-temporal interactions with a graph attention module that leverages transformers, in order to output all the plausible multimodal trajectories through generative learning. From our experimental inferences, it is apparent that our proposed model consistently eclipses the state-of-the-art (SOTA) models, on an average, with a noticeable decrease in error rate of 40.20% in average displacement error (ADE) and 37.16% in final displacement error (FDE).</p>