Multi-adversarial Adaptive Transformers for Joint Multi-agent Trajectory Prediction

Multi-agent trajectory prediction is of vital importance for autonomous driving and robotic systems, particularly in situations where frequent interaction happens. Existing methods essentially suppose that training and testing data are drawn from the identical distribution, while ignoring the potential domain discrepancy. This is not hold in practice and results in inevitable performance degradation. To alleviate the problem of domain discrepancy, we propose a novel multi-adversarial adaptive transformers framework, which jointly conducts multi-agent trajectory prediction and domain adaptation in a unified framework. Specifically, the framework consists of a simple but effective transformer-based encoder-decoder architecture and three domain adaptation components. The former generates multi-modal trajectories of multi-agents simultaneously, and the latter reduces the domain disparity from different aspects: the temporal aspect, the social aspect, and the contextual aspect. The three domain adaptation components are implemented by learning a domain classifier in an adversarial training manner, respectively. By this way, domain-invariant feature representations are learned and domain discrepancies will be better alleviated. Practical and challenging experiments are conducted cross multiple domains, and the results demonstrate the effectiveness of our proposed framework for robust multi-agent trajectory prediction.