Towards Human Motion Prediction for Collaborative Robotics with Text-to-Motion Data Generation

Human Motion Prediction (HMP) in Collaborative Robotics enables proactive adaptation of robot behavior, improving efficiency and safety. However, the available data from human-robot interaction scenarios is scarce and collecting new motion samples is a challenging and expensive process. In this work, we propose a novel approach for HMP that exploits synthetic human motion sequences to mitigate the need for real-world data collection. Our method consists of generating a large and diverse collection of human motions that are unified into a synthetic dataset for model pre-training. The pre-trained model is then fine-tuned on the small available dataset of real motion sequences from a target context. For a comprehensive evaluation of the effectiveness of our approach, we generate three synthetic datasets of increasing size and evaluate their impact on three HMP models with different architectures. We then compare the prediction accuracy of our approach with the same models trained exclusively on the target dataset. The results showed a strong improvement in HMP given by pre-training on synthetic sequences. The increase in prediction accuracy was consistent across all the models, reaching up to 19.49% error reduction. This outcome highlights the applicability of synthetic motion data in HMP independent of the specific model architecture. By using synthetic data for pre-training, our approach enables more effective model adaptation to real-world applications, making it a promising solution for enhancing human-robot collaboration.