AI-based automated RULA and REBA assessment using video analysis for ergonomics risk evaluation

Introduction

Musculoskeletal disorders (MSDs) represent one of the most prevalent and costly occupational health problems worldwide. According to the World Health Organization, MSDs are the leading contributor to global disability, with work-related MSDs substantially impacting productivity, worker well-being, and healthcare expenditure. In manufacturing and service industries, prolonged adoption of non-neutral postures, repetitive motions, and awkward working positions significantly elevate the risk of developing upper limb, neck, and back disorders. Ergonomic risk assessment tools such as the Rapid Upper Limb Assessment (RULA) and Rapid Entire Body Assessment (REBA) are widely validated methods for quantifying postural risk in occupational settings. RULA focuses on the upper extremities, wrists, forearms, upper arms, and neck and assigns a risk score to guide intervention priorities. REBA extends this to include whole-body postures, incorporating trunk, legs, and load/force requirements, making it particularly suitable for dynamic, unpredictable activities. Despite their effectiveness, traditional RULA and REBA assessments are manual, observer-dependent, and resourceintensive. An ergonomist must directly observe workers, record posture data, and score each body segment, a process that is impractical for large-scale or continuous monitoring. Inter-rater variability is also a recognized limitation, where two trained observers may assign differing scores for the same task. The emergence of AI-driven computer vision technologies offers a promising solution to these challenges. Pose estimation algorithms, such as those used in platforms like TuMeke, can automatically detect skeletal joint positions from standard video footage and compute ergonomic scores in near real-time. This enables scalable, objective, and costefficient ergonomic screening across multiple workstations and shifts. This paper evaluates the practical application of computer vision-based RULA and REBA assessment in a real-world manufacturing environment. It presents findings from a study conducted at a manufacturing facility in Bangalore, India, assesses the risk levels identified across diverse workplace tasks, and critically examines the strengths and limitations of AI-based ergonomic tools as part of an integrated occupational health strategy.

Objectives

The specific objectives of this study were:

• To evaluate the feasibility of computer vision-based automated RULA and REBA assessment in industrial and office workstations. • To identify ergonomic risk levels across different workplace tasks using automated posture analysis. • To examine the opportunities and limitations of AIdriven ergonomic assessment as a decision-support tool in occupational health practice.

Review of Literature 3.1 Traditional Ergonomic Assessment Methods

RULA was developed by McAtamney and Corlett (1993) [5] as an observational method to assess the exposure of individual workers to ergonomic risk factors associated with upper limb work-related disorders. It evaluates postures of the upper arm, lower arm, wrist, neck, trunk, and legs, producing a grand score from 1-7 that indicates urgency for intervention. REBA, introduced by Hignett and McAtamney (2000) [3] , extended this approach to the whole body and incorporated dynamic load handling, becoming the preferred tool for assessing jobs involving unpredictable postures such as those in healthcare and logistics. Both tools have been extensively validated and are recommended by occupational safety standards globally. However, they are inherently observational, relying on trained ergonomists to capture task postures at representative moments, which introduces subjectivity and limits their scalability. Studies have noted inter-rater reliability challenges, with intraclass correlation coefficients varying widely across assessors (Dempsey & Mathiassen, 2006) [2] .

AI and Computer Vision in Ergonomics

The integration of machine learning and computer vision into ergonomic assessment has been an active research area since the mid-2010s. Deep learning-based pose estimation models, such as OpenPose (Cao et al., 2017) [1] , AlphaPose, and MediaPipe, can detect 2D or 3D skeletal joint coordinates from video footage with high accuracy, enabling automated computation of joint angles and postural classifications.

Several studies have demonstrated the validity of automated RULA and REBA scoring against manual expert assessment. Vignais et al. (2013) [6] showed that inertial measurement unit (IMU)-based systems could produce RULA scores comparable to observational methods. More recently, RGB camera-based systems have achieved similar or superior reliability. Platforms such as TuMeke leverage these technologies to provide real-time ergonomic risk feedback from standard video recordings, making them practically deployable in industrial settings. Despite the promise of automation, a number of limitations persist. Occlusion of body segments, variation in camera angle, clothing effects on joint detection, and the inability to assess contact forces or tool use represent ongoing challenges. Moreover, automated systems may not capture the full temporal context of a task, cumulative exposure over a shift, which is an important determinant of MSD risk (Mathiassen, 2006) [4] .

Materials and Methods

Sudy Design and Setting

This study employed a cross-sectional observational design.

The investigation was conducted at two operational units of a manufacturing facility located in Bangalore, Karnataka, India. The facility produces electronics components and operates under standard shift conditions. Work areas assessed included receiving, incoming inspection, testing, debug and repair, shipment packing, warehousing (stores), final inspection and quality assurance, and office workstations, providing a comprehensive representation of both industrial and sedentary occupational tasks.

Participants

A total of 60 workers participated in the study (n = 60). Workers were purposively selected to represent all major task categories across both study units. Participation was voluntary, and all participants provided informed consent prior to data collection. Ethical clearance was obtained from the institutional review board prior to the commencement of the study.

Instrumentation

Two primary tools were employed for automated postural assessment:

TuMeke (Computer Vision-Based Posture Estimation

System): TuMeke uses AI-powered skeletal detection to analyse video recordings of workers performing their tasks.

The system automatically identifies joint positions frameby-frame, computes joint angles, and generates RULA and REBA scores in accordance with established scoring methodologies. It provides risk classification outputs on a per-task basis, enabling rapid ergonomic screening without the need for manual observation.

APECS (AI Posture Evaluation and Correction System):

APECS was used specifically for head and neck postural assessment. It measures three key angular parameters:

Discussion 6.1 Opportunities of AI-Based Ergonomic Assessment

The results of this study demonstrate that computer visionbased RULA and REBA assessment can effectively identify non-neutral postures and ergonomic risk levels across a diverse range of industrial and office tasks. The automated system enabled consistent and rapid screening of 60 workers across multiple work areas a scale that would have required substantially more time and personnel using traditional observational methods.

A key advantage of AI-driven assessment is its objectivity. Unlike manual RULA and REBA evaluations, which are susceptible to observer bias and inter-rater variability, automated pose estimation applies consistent algorithmic criteria to all observations. This is particularly valuable in large organisations seeking to standardise ergonomic monitoring programmes or conduct longitudinal comparisons of risk levels over time. The integration of APECS for head and neck posture analysis adds a valuable dimension not typically captured by standard RULA and REBA assessments. By quantifying CHA, CVA, and SSP parameters, APECS provides nuanced data on postural deviations in the cervical and shoulder region, an area of growing concern given the rise of screenbased and precision assembly work. Furthermore, video-based assessment is non-intrusive and can be conducted without disrupting normal work operations. Workers do not need to wear sensors or markers, and recordings can be analysed retrospectively, offering flexibility in data collection scheduling.

Limitations of Automated Assessment

Despite these advantages, several important limitations were identified. First, automated systems may struggle with tasks involving occlusion, where body segments are obscured by equipment, other workers, or the worker's own clothing. In such cases, pose estimation accuracy decreases, potentially producing erroneous or incomplete scores. Second, automated RULA and REBA scores represent instantaneous postural snapshots or averages over recorded intervals, but do not fully capture the cumulative exposure experienced by a worker over the course of a full shift. Cumulative biomechanical load, a key determinant of MSD risk, requires more sophisticated exposure assessment models that account for the frequency, duration, and www.extensionjournal.com magnitude of postural deviations across the working day. Third, task context and work organisation factors, such as production pace, task variety, recovery opportunities, and psychosocial stressors are not captured by posture estimation algorithms. These factors are integral to comprehensive ergonomic risk assessment and can substantially modify the risk implied by a given postural score. Fourth, while the systems used in this study provide risk scores, the interpretation of these scores and the formulation of effective ergonomic interventions requires domain expertise. Automated tools identify where the risk exists, but they cannot independently determine why the risk exists or what the most appropriate intervention strategy would be. Over-reliance on automated outputs without expert interpretation may lead to inadequate or misdirected interventions.

AI as a Decision-Support Tool

Given these opportunities and limitations, this study affirms that AI-based ergonomic assessment is most effectively positioned as a decision-support tool rather than a standalone assessment method. In this role, automated RULA and REBA screening can serve as the first tier of a multi-stage ergonomic evaluation process, efficiently identifying high-risk tasks and workstations for priority attention, thereby enabling ergonomists to direct their expertise where it is most needed. Integration with expert human factors evaluation is essential to contextualise automated findings, account for cumulative exposure and work organisational factors, and develop targeted interventions. This hybrid model leverages the scalability and consistency of AI while preserving the interpretive depth of human expertise.

Conclusions

This study evaluated the feasibility and utility of computer vision-based automated RULA and REBA assessment in a real-world manufacturing environment. The principal conclusions are as follows:

• AI-driven posture analysis using TuMeke effectively identified non-neutral postures across a diverse range of industrial and office tasks in a manufacturing facility. • Most tasks were classified within the low to moderate ergonomic risk range, with higher risks observed in tasks involving manual material handling, sustained forward trunk flexion, and precision assembly work. These findings support the adoption of AI-based ergonomic tools as part of proactive occupational health management programmes, provided they are deployed within a robust framework that includes expert oversight, periodic validation, and integration with broader workplace health initiatives.

Recommendations

Based on the findings of this study, the following recommendations are proposed for organisations implementing ergonomic assessment programmes:

• Use computer vision-based RULA and REBA tools for initial ergonomic risk screening across all workstations, with priority given to tasks identified as medium to high risk.

Acknowledgement

AcknowledgementsThe authors gratefully acknowledge EHS Guru Sustainable Solutions Pvt. Ltd. for their assistance and support in granting permission and access to conduct this ergonomic assessment across the facilities, enabling the successful completion of the study. The authors also extend their appreciation to the management and workers of the participating manufacturing facility for their cooperation and willingness to participate in the research.

References

1. Cao Z, Simon T, Wei SE, Sheikh Y. Realtime multiperson 2D pose estimation using part affinity fields. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR); 2017. p. 72917299.