TATIC: Task-Aware Temporal Learning for Human Intent Inference from Physical Corrections in Human-Robot Collaboration

With the rise of human-robot collaboration, robots need to quickly adapt to dynamic task constraints and evolving human intent in shared workspaces. Traditional physical human-robot interaction methods primarily rely on physical corrections for trajectory guidance but struggle to infer task-level semantic intent. Recent advances in semantic planning and robot control have been driven by large language models and vision-language models, but these methods mainly rely on vision and language inputs, lacking mechanisms to interpret physical feedback.

In dynamic human-robot collaboration environments, robots must quickly adapt to changing task constraints and human intent. Existing methods fall short in handling physical feedback, unable to effectively decode semantic intent and motion corrections from physical interactions.

The TATIC framework combines torque-based contact force estimation with a task-aware Temporal Convolutional Network to infer discrete task-level intent and estimate continuous motion-level parameters from brief physical corrections. Task-aligned feature canonicalization ensures robust generalization across diverse layouts, while an intent-driven adaptation scheme translates inferred human intent into robot motion adaptations.

• �� Utilize torque estimation for contact force estimation, supporting brief physical corrections. • Employ a task-aware Temporal Convolutional Network (TCN) to infer discrete task-level intent and continuous motion-level parameters. • Ensure robust generalization across diverse layouts through task-aligned feature canonicalization. • Translate inferred human intent into robot motion adaptations through an intent-driven adaptation scheme.

Experiments were conducted on a 7-DoF manipulator, collecting 500 in-distribution and 250 out-of-distribution episodes. The experimental design included intent recognition tasks and hardware validation, using Macro-F1 score as the primary evaluation metric. Feature ablation studies evaluated the impact of different feature sets on model performance.

Experimental results show TATIC achieves a 0.904 Macro-F1 score in intent recognition, demonstrating its efficiency in multi-task environments. Hardware validation successfully applied TATIC in collaborative disassembly tasks, showcasing its applicability in real-world scenarios. Feature ablation studies indicate that using the full feature set significantly improves model performance, particularly in target selection and direction regression tasks.

TATIC can be applied in dynamic environments requiring rapid human intent response, such as manufacturing, disassembly, and healthcare assistance. Its low-latency interaction and efficient intent recognition capabilities make it highly applicable in these fields.

Despite TATIC's strong performance, challenges remain in complex multi-contact scenarios. The model assumes a single contact point, which may struggle in multi-contact scenarios. Additionally, prolonged physical interactions may lead to operator fatigue, and the model's performance under extreme environmental conditions remains unverified.