Probabilistic Human Intent Prediction for Mobile Manipulation: An Evaluation with Human-Inspired Constraints

Abstract We present GUIDER (Global User Intent Dual-phase Estimation for Robots), a dual-phase probabilistic framework for intent inference in mobile manipulation that operates without predefined goals. A Synergy Map fuses motion evidence with an occupancy grid to rank likely interaction areas during navigation. After arrival, perception merges U $$^{2}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mmultiscripts> <mml:mrow/> <mml:mrow/> <mml:mn>2</mml:mn> </mml:mmultiscripts> </mml:math> -Net and FastSAM saliency with three geometric grasp-feasibility tests; an end-effector kinematics-aware update then evolves object probabilities in real time. In 100 teleoperation trials (20 participants $$\times $$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mo>×</mml:mo> </mml:math> 5 tasks) in Isaac Sim, GUIDER outperformed baselines. During navigation, median stability was 100% across tasks (BOIR, the baseline, had an overall median of 89.85%), with large gains under redirection (BOIR 59.67–63.49% in T2/T5). During manipulation, median stability was 100% in all tasks, while Trajectron (manipulation baseline) dropped to 62.68% for tool grasping (T4). GUIDER yielded earlier confident object predictions in geometry-constrained settings (T5: 20.31 s remaining vs 3.89 s). Ablations confirm the need for the multi-horizon synergy map, the grasp-feasibility checks, and temporal end-effector probability evolution. GUIDER provides a unified probabilistic backbone spanning base and arm, supporting future variable-autonomy controllers.