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

A research team led by Jiurun Song, Xiao Liang, and Minghui Zheng has introduced TATIC (Task-Aware Temporal Learning for Human Intent Inference), a novel AI framework designed to revolutionize physical human-robot collaboration (HRC). The core challenge TATIC addresses is the semantic gap in communication: while a human can physically nudge a robot's arm to correct its motion, existing systems struggle to understand the higher-level *task intent* behind that brief contact. Traditional physical interaction methods see the push but miss the purpose, and modern vision/language models lack the ability to interpret this tactile channel. TATIC bridges this by combining real-time torque-based contact force estimation with a specialized Temporal Convolutional Network (TCN) that is explicitly "task-aware."

This architecture allows TATIC to perform a dual inference: it simultaneously recognizes the discrete task a human wants to accomplish (like "unscrew" or "lift") and estimates the continuous motion parameters (like direction and force) from the same physical interaction. A key innovation is "task-aligned feature canonicalization," which helps the system generalize robustly across different workspace layouts and scenarios. In experiments, TATIC demonstrated a high intent recognition accuracy, scoring a 0.904 Macro-F1 score. The framework was successfully validated on hardware in a collaborative disassembly task, where an intent-driven adaptation scheme allowed the robot to immediately adjust its movements based on the inferred human goal, moving beyond simple trajectory following to true task-level understanding.