Understanding human motion behavior is critical for autonomous moving\nplatforms (like self-driving cars and social robots) if they are to navigate\nhuman-centric environments. This is challenging because human motion is\ninherently multimodal: given a history of human motion paths, there are many\nsocially plausible ways that people could move in the future. We tackle this\nproblem by combining tools from sequence prediction and generative adversarial\nnetworks: a recurrent sequence-to-sequence model observes motion histories and\npredicts future behavior, using a novel pooling mechanism to aggregate\ninformation across people. We predict socially plausible futures by training\nadversarially against a recurrent discriminator, and encourage diverse\npredictions with a novel variety loss. Through experiments on several datasets\nwe demonstrate that our approach outperforms prior work in terms of accuracy,\nvariety, collision avoidance, and computational complexity.\n