HemiBot: design and technical validation of an adaptive Robotic coaching system for Unilateral Spatial Neglect rehabilitation

Abstract

Robot-mediated neuro-motor rehabilitation offers intensive, repeatable, and objectively monitored therapy, yet existing systems rarely integrate automated session management, real-time per-repetition scoring, and data-driven adaptive difficulty progression within a single non-contact architecture. This work presents HemiBot (Hemineglect roBot), a ROS 2 software stack implementing a fully automated coaching pipeline for semicircular arm-pointing exercises on the TIAGo mobile manipulator (PAL Robotics), targeting post-stroke Unilateral Spatial Neglect (USN), in which patients fail to orient toward or interact with stimuli in the contralesional hemispace. The robot’s end effector moves into the neglected hemispace as an exogenous attentional cue while the participant actively reaches toward it with the affected arm. The system extends a skeletal tracking framework with a Behavior Tree orchestrating exercise execution and per-repetition metric computation, a Coach node performing session-level adaptive difficulty progression, Italian empathetic text-to-speech feedback, and a scoreboard for therapist supervision. The scoring metric combines arm-alignment angle and hand–end-effector distance in a logical AND gate, yielding a normalized 0–100 score partitioned into three performance bands (Fail <25, Acceptable 25–75, Great ≥75). A participant is promoted when the mean session score exceeds 75 and demoted when the fail ratio reaches 0.50. Sixteen healthy participants (mean age 54.5 ± 6.8 years) were divided into two groups of eight: a high-performance group performing all repetitions correctly, and a low-performance group following a controlled error pattern designed to trigger demotion. Kinematic time series from the robot’s onboard RGB-D camera were compared against a stereophotogrammetric gold standard (BTS SMART-DX EVO). Scoring concordance between TIAGo- and mocap-derived classifications reached 100% (Cohen’s κ = 1.000) across all 120 repetitions, with no cross-band errors between Fail and Great. The executed arc trajectory showed near-perfect agreement with the commanded geometry (Pearson r > 0.99). Adaptive behavior was correct for all 16 participants. These findings verify that the onboard measurement chain reproduces gold-standard band assignments under the tested conditions and that the session-level decision pipeline operates end-to-end as specified. Validation of the scoring thresholds under naturally occurring, near-boundary performance remains necessary before clinical use.