Frailty is a key predictor of adverse health outcomes in older adults, and early detection through standardised assessments such as the Short Physical Performance Battery (SPPB) and the Timed Up and Go (TUG) can help prevent falls and functional decline. However, these assessments are time-consuming for healthcare professionals and typically rely on coarse outcome measures. We present a robotic framework that autonomously administers SPPB and TUG assessments with older adults in a clinical environment. The system uses a social robot (Temi) with an external stereoscopic camera for markerless motion capture. A Behaviour Tree controls the assessment workflow, guiding participants through each test with verbal and visual instructions while monitoring their movements in real time. We validated our framework in a longitudinal study with 57 older adults (mean age 84.1 years) over 116 sessions. We assessed agreement between robot-derived and therapist-administered measurements for the SPPB total score, individual test scores, and completion times. We also evaluated the framework’s usability using the System Usability Scale. Our results demonstrate strong correlation (Pearson’s r = 0.99) and substantial agreement (Lin’s CCC = 0.98) between robot and therapist measurements for the SPPB total score. The usability assessment yielded an average SUS score of 74.0, indicating good usability. These findings validate the feasibility of deploying autonomous robotic systems for standardised frailty assessment in clinical settings.
Clinical Needs and Co-Design
The robotic framework was developed through a user-centred design process with healthcare professionals from the Intermediate Care Hospital Parc Sanitari Pere Virgili. A multidisciplinary co-design workshop involving geriatricians, rehabilitation physicians, physiotherapists, occupational therapists, nurses, and nursing assistants identified three recurring challenges: standardised assessments require dedicated time from healthcare professionals, clinical evaluations rely primarily on coarse outcome measures while informative biomechanical indicators are rarely collected, and maintaining standardised test administration while reducing workload is essential.
Participants envisioned a system capable of autonomously guiding older adults through standardised protocols, collecting objective measurements, and allowing professionals to focus on interpreting results. The physical component of frailty was prioritised as the most suitable starting point due to its standardised procedures and routine clinical use.
Platform Selection and Initial Prototype
The clinical requirements guided implementation of the two most widely adopted tests: the Short Physical Performance Battery (SPPB) and the Timed Up and Go (TUG). The Temi social robot was selected as the interaction platform, integrating a touchscreen, speakers, microphones, and autonomous mobile base. Because onboard computational resources were insufficient for real-time vision-based skeleton tracking, a stereoscopic ZED2i camera connected to an external GPU-equipped computer was integrated into the system.
Iterative Refinement and Pilot Deployment
The initial prototype underwent several refinement cycles with a physiotherapist, occupational therapist, and neuropsychologist. A pilot deployment with four older adults in the physiotherapy laboratory led to three final modifications. The stereoscopic camera was repositioned perpendicular to the walking direction to improve skeleton tracking accuracy during Chair Stand and TUG assessments. Standing Balance instructions were revised to encourage participants to look forward instead of toward the robot’s display. Verbal feedback during Chair Stand repetitions was removed to avoid influencing performance. An occupational therapist supervised all evaluations for safety.
Robotic Framework for Autonomous Frailty Assessment
The complete framework integrates perception, decision-making, and interaction modules to autonomously conduct frailty and fall-risk assessments.
Perception Module
A stereoscopic depth camera (ZED2i) observes participants at distances up to 10 meters, capturing both static tests and dynamic walking tasks. Human pose estimation uses the ZED SDK skeleton tracking module, estimating 3D positions of 38 body keypoints representing major joints including head, torso, pelvis, hips, knees, ankles, and feet. Detected 3D joint coordinates are continuously streamed and stored in a temporal buffer for movement event detection and biomechanical metric computation. Noise in the skeleton estimation is reduced through filtering before further processing.
Decision-Making Module
A Behaviour Tree (BT) controls the assessment workflow. The system first checks hardware connections, then introduces itself to the participant and explains the assessment purpose. The BT guides participants through Standing Balance, Gait Speed, Chair Stand, and Timed Up and Go tests in sequence. Each test is implemented as an action node that explains the procedure, monitors movements once the participant confirms readiness, and detects relevant events such as movement start and completion. The BT handles test termination conditions appropriately, such as ending Standing Balance when imbalance is detected according to the SPPB protocol.
Interaction Module
Participants interacted with the robot exclusively through touch-based inputs on the robot’s touchscreen display. Speech recognition was not implemented due to reduced accuracy with older adults’ age-related vocal changes.
Experimental Validation
Study Design and Participants
We conducted a longitudinal study at the Parc Sanitari Pere Virgili Intermediate Care Hospital. Fifty-seven older adults (mean age 84.1 years, standard deviation 8.9) participated in 116 sessions. Inclusion criteria were age 65 years or older and ability to walk independently with or without walking aids. Exclusion criteria included severe cognitive impairment, acute medical conditions, and inability to follow simple instructions.
Data Collection Protocol
Each session involved both a therapist-administered and a robot-administered frailty assessment in randomised order. The therapist conducted assessments using standard clinical procedures with a stopwatch. The robot administered the same assessments autonomously using the framework described above. Both sessions were recorded with video cameras for ground truth verification.
Outcome Measures
Primary outcome measures included SPPB total score (range 0-12), individual component scores (Standing Balance 0-4, Gait Speed 0-4, Chair Stand 0-4), and completion times for each component. Usability was assessed using the System Usability Scale (SUS), yielding a score from 0 to 100.
Results
Agreement Between Robot and Therapist Measurements
Strong correlation was observed between robot and therapist measurements. The SPPB total score showed Pearson’s r = 0.99 and Lin’s concordance correlation coefficient (CCC) = 0.98. Gait speed scores showed Pearson’s r = 0.97 and CCC = 0.96. Sit-to-stand scores showed Pearson’s r = 0.93 and CCC = 0.92. Standing balance scores showed Pearson’s r = 0.93 and CCC = 0.90.
Completion Time Agreement
Completion time measurements showed strong agreement across all tests. Gait speed time measurements yielded mean absolute error of 0.21 seconds (10.4% relative error). Sit-to-stand time measurements showed mean absolute error of 0.63 seconds (12.1% relative error). Standing balance time measurements achieved mean absolute error of 0.40 seconds (4.9% relative error).
Usability Assessment
The System Usability Scale assessment yielded an average score of 74.0 (standard deviation 12.3), indicating good usability. This score falls within the range of acceptable to excellent according to standard SUS interpretation guidelines.
Discussion
The results demonstrate that an autonomous robotic system can reliably administer standardised frailty assessments in a real clinical environment with older adults. The strong agreement between robot and therapist measurements validates the framework’s measurement accuracy. The robot successfully extracted additional biomechanical metrics beyond conventional clinical scores, addressing the clinical need for richer mobility indicators.
The usability assessment indicates that older adults found the system acceptable and easy to use despite their advanced age (mean 84.1 years). This finding suggests that autonomous robotic assessment can be integrated into routine clinical practice without imposing additional burden on patients.
Limitations
The study was conducted in a single hospital with a relatively homogeneous participant population. The robot required external computing hardware and a dedicated camera setup, which may limit deployment flexibility. The framework currently assesses only the physical component of frailty, while comprehensive assessment also requires cognitive, social, and nutritional components.
Conclusion
This work validates the feasibility of deploying autonomous robotic systems for standardised frailty assessment in real clinical settings. The robotic framework demonstrates strong measurement agreement with therapist-administered assessments and acceptable usability for older adults. These findings support further development toward routine autonomous frailty screening that could reduce healthcare professional workload while providing richer, objective measurements of functional performance.
Frequently Asked Questions
What frailty assessments can the robot perform autonomously? The robot autonomously administers the Short Physical Performance Battery (SPPB) including Standing Balance, Gait Speed, and Chair Stand tests, as well as the Timed Up and Go (TUG) test.
How accurate are the robot measurements compared to a human therapist? Robot measurements show strong correlation with therapist measurements, with SPPB total score correlation of r = 0.99 and Lin’s concordance correlation coefficient of 0.98.
Did older adults find the robot easy to use? Yes, the average System Usability Scale (SUS) score was 74.0 out of 100, indicating good usability for the older adult population with mean age 84.1 years.
What sensing technology does the robot use to measure movement? The system uses a stereoscopic ZED2i depth camera with real-time 3D skeleton tracking of 38 body keypoints, requiring no wearable sensors or markers on the participant.