The demand for real-time perception in complex robotic systems operating in dynamic environments has motivated the development of architectures capable of solving multiple learning tasks simultaneously. However, current Multi-Task Learning approaches often suffer from poor task balancing, lack of inter-task regularization, and static loss weighting. This work introduces a multi-task framework based on Extended Long Short-Term Memory (xLSTM) to jointly predict the motion of 3D bounding boxes, semantic classes, 3D velocity, and categorical dynamic behavior of objects. The framework adopts an encoder–multi-head architecture with shared temporal representations and task-specific heads. Two auxiliary tasks, velocity regression and dynamic state classification, are derived from physical approximations and incorporated to guide training. A homoscedastic uncertainty-based loss weighting strategy dynamically adjusts task influence during optimization. Quantitative results on the KITTI benchmark show that the proposed framework achieves lower RMSE for motion estimation and higher F1-scores for classification tasks compared to baselines. Auxiliary tasks improve convergence and coherence, while uncertainty weighting enhances training stability. This architecture offers a scalable and interpretable solution for spatiotemporal modeling and has the potential to benefit downstream applications in robotics and intelligent monitoring systems.