SimpliHuMoN: Simplifying Human Motion Prediction

ICLR 2026 Conference Submission21869 Authors

Keywords: Human Motion Prediction, Generalist Model, Pose Prediction, Trajectory Prediction

TL;DR: We show a simple Transformer model achieves SOTA results across all motion prediction tasks (pose, trajectory, combined), challenging the trend toward complex, specialized architectures.

Abstract:

Human motion prediction combines the tasks of trajectory forecasting, human pose prediction, and possibly also multi-person modeling.
For each of the three tasks, specialized, sophisticated models have been developed due to the complexity and uncertainty of human motion. While compelling for each task, combining these models for holistic human motion prediction is non-trivial. Conversely, holistic human motion prediction methods, which have been introduced recently, have struggled to compete on established benchmarks for individual tasks. To address this dichotomy, we study a simple yet effective model for human motion prediction based on a transformer architecture. The model employs a stack of self-attention modules to effectively capture both spatial dependencies within a pose and temporal relationships across a motion sequence. This simple, streamlined, end-to-end model is sufficiently versatile to handle pose-only, trajectory-only, and combined prediction tasks without task-specific modifications. We demonstrate that our approach achieves state-of-the-art results across all tasks through extensive experiments on a wide range of benchmark datasets, including Human3.6M, AMASS, ETH-UCY, and 3DPW. Our results challenge the prevailing notion that architectural complexity is a prerequisite for achieving accuracy and generality in human motion prediction. Code will be released.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 21869