Time-Series Prediction for Intent Aware Robot Learning

Abstract

The study of time-series prediction and intent modeling forms a fundamental component of intelligent and adaptive systems in robotics and human–machine interaction. Accurately modeling temporal dependencies and anticipating user intent are essential for achieving seamless collaboration, safety, and autonomy in systems such as collaborative robots, eXtended Reality (XR) interfaces, and virtual pilot assistance platforms. Despite substantial advances in modeling and control, existing approaches often struggle to represent the stochastic, nonlinear, and context-sensitive characteristics of human behavior in dynamic environments. This thesis addresses these challenges by developing and investigating probabilistic and learning-based frameworks for time-series prediction across diverse interaction scenarios. The research begins with probabilistic formulations for intent prediction during human– robot handover. A framework based on Maximum Entropy Deep Inverse Reinforcement Learning (MEDIRL) was developed to infer human intent from partial hand trajectories by modeling motion as a reward-driven process. This approach was extended to a multimodal setting that integrates hand motion and eye-gaze, thereby enhancing robustness and improving early intent recognition. The inclusion of gaze information enabled the model to capture attention-driven behavior and reduce ambiguity in predicting users’ intended targets during human-robot handover. Building on these foundations, the thesis advances to multi-agent time-series prediction for robot-assisted manufacturing. A Feature-Based Bayesian Interaction Primitive (FBIP) formulation was proposed to extend classical Bayesian Interaction Primitives by embedding task-relevant feature functions within the probabilistic model. This framework enabled the prediction of coordinated motion patterns between human and robot. The probabilistic representation captured inter-agent dependencies effectively. To address long-horizon temporal forecasting, the thesis introduces BiPTraP (Bayesian Interaction Primitives with Transformer for Predicting Time-Series Data)—a hybrid model that combines Bayesian state estimation with Transformer-based self-attention. Trajectories were represented in a basis space, updated probabilistically using ensemble Kalman filtering, and encoded as patches for Transformer-based temporal reasoning. BiPTraP demonstrated superior accuracy and stability in predicting non-stationary time-series data, outperforming conventional recurrent and probabilistic baselines. The final part of the thesis explores applications of time-series prediction in two contrasting domains. In XR interaction, a Sampling-based Maximum Entropy IRL (SMEIRL) framework was introduced for rapid target prediction during virtual and mixed-reality pointing tasks, achieving high prediction accuracy with efficient sampling. In aviation, state-ofthe- art sequence models—LSTM, Mamba, Jamba, and PatchTST—were benchmarked for pilot control input prediction using flight simulator data. The PatchTST achieved the lowest prediction error, highlighting its potential for developing intelligent virtual pilot assistance systems. Overall, the thesis presents a progression from probabilistic intent inference to scalable Transformer-based architectures for time-series prediction. The proposed methods demonstrate early intent recognition, multimodal fusion, and long-horizon forecasting capabilities across human–robot, XR, and flight control applications, contributing to the broader goal of enabling predictive, adaptive, and intent-aware autonomous systems.