Multimodal Trajectory Generation for Robotic Motion Planning Using Transformer-Based Fusion and Adversarial Learning

In Human–Robot Interaction (HRI), generating robot trajectories that accurately reflect user intentions while ensuring physical realism remains challenging, especially in unstructured environments. In this study, we develop a multimodal framework that integrates symbolic task reasoning with continuous trajectory generation. The approach employs transformer models and adversarial training to map high-level intent to robotic motion. Information from multiple data sources, such as voice traits, hand and body keypoints, visual observations, and recorded paths, is integrated simultaneously. These signals are mapped into a shared representation that supports interpretable reasoning while enabling smooth and realistic motion generation. Based on this design, two different learning strategies are investigated. In the first step, grammar-constrained Linear Temporal Logic (LTL) expressions are created from multimodal human inputs. These expressions are subsequently decoded into robot trajectories. The second method generates trajectories directly from symbolic intent and linguistic data, bypassing an intermediate logical representation. Transformer encoders combine multiple types of information, and autoregressive transformer decoders generate motion sequences. Adding smoothness and speed limits during training increases the likelihood of physical feasibility. To improve the realism and stability of the generated trajectories during training, an adversarial discriminator is also included to guide them toward the distribution of actual robot motion. Tests on the NATSGLD dataset indicate that the complete system exhibits stable training behaviour and performance. In normalised coordinates, the logic-based pipeline has an Average Displacement Error (ADE) of 0.040 and a Final Displacement Error (FDE) of 0.036. The adversarial generator makes substantially more progress, reducing ADE to 0.021 and FDE to 0.018. Visual examination confirms that the generated trajectories closely align with observed motion patterns while preserving smooth temporal dynamics.