Enhancing human motion prediction through integration of spatial context and object awareness

Improving human-robot interaction is a critical factor in increasing the safety, productivity, and adaptability of collaborative robotics. A key challenge in this area is to ensure seamless and safe human-machine interaction without compromising production efficiency. A major limitation is the robot's inability to effectively understand and predict the operator's movements and intentions. Enhancing a robot's awareness of its environment, including the presence and actions of human operators, can significantly improve both performance and safety. This thesis addresses this challenge by focusing on Human Motion Prediction (HMP), a technique that allows the anticipation of human movements based on past poses. By predicting future poses, robots can proactively adjust their trajectories, resulting in smoother, safer and more efficient task execution. Enhanced environmental and human awareness has the potential to transform collaborative workflows by achieving a critical balance between operational performance and safety. This thesis presents a novel machine learning architecture that improves prediction accuracy by incorporating spatial features. Using an encoder-decoder framework, the model combines past operator movements with spatial semantic information about surrounding objects and obstacles via a graph-based spatial encoder. This approach uniquely integrates the dynamics of human motion with the context described by relevant features of the environment. By taking into account the positions and volumes of objects of interest and obstacles, the model can generate realistic and environmentally consistent predictions. The proposed model was trained and validated on datasets combining human joint rotation data with bounding box representations of objects and obstacles. Using the GRAB (GRasping Actions with Bodies) dataset for initial experiments, the model showed significant improvements in prediction accuracy compared to baselines without spatial semantics. The inclusion of the KIT Whole-Body Human Motion Database expanded the dataset with additional multi-object scenarios, increasing the variety of input data and improving the model's generalization and adaptability to real-world applications. The architecture, now capable of handling a variable number of input objects, provides increased flexibility and suitability for use in dynamic, unstructured environments, bridging the gap between advanced HMP techniques and practical real-world scenarios. The results of this work highlight the impact of integrating spatial semantic information into HMP and represent a significant step toward safer and more efficient human-robot collaboration. By advancing the integration of contextual awareness into predictive models, this research opens new horizons for dynamic, real-world applications where accurate anticipation of human motion is critical. The proposed approach lays a solid foundation for future innovations with implications beyond robotics, shaping the future of human-machine synergy in complex environments.