AI Skeleton Tracking and Manual Ergonomics Software Face Off on the Factory Floor

A worker in a vehicle maintenance shop leans into a passenger car to remove a rear door, and in that ordinary motion lies a question that has troubled occupational health researchers for decades: how much physical load is that task really placing on the body? A new study from the Technical University of Košice in Slovakia offers one of the most concrete answers yet, by pitting two very different digital ergonomics platforms against each other on exactly the same footage. Matej Cichý, Darina Dupláková and Viktória Tutokyová recorded the sequential removal of a rear car door under real workshop conditions and then fed the identical video into two assessment systems running in parallel: the LEA mobile application, version 2.3.1, which relies on automated artificial intelligence skeleton tracking, and the Timer Pro Professional software, version 10.1, which depends on an analyst manually parameterizing each posture. The comparison, published in Mobile Networks and Applications, is one of the first head-to-head tests of automated versus conventional digital ergonomics conducted dynamically, in situ, rather than in a laboratory.

The stakes of this kind of work are far from academic. Work-related musculoskeletal disorders remain among the largest categories of occupational disease worldwide, and the scientific literature the authors draw upon links awkward postures, repetitive mechanical exposure and lifting to chronic low back pain and upper limb disorders. In the automotive repair and manufacturing sectors specifically, recent systematic reviews and meta-analyses report strikingly high prevalence of these disorders among workers, driven by constrained postures, hand force exertion and non-neutral trunk positions. Traditional observational tools such as RULA, the Rapid Upper Limb Assessment method introduced by McAtamney and Corlett in 1993, and REBA, the Rapid Entire Body Assessment proposed by Hignett and McAtamney in 2000, were designed for a human assessor with a clipboard or a spreadsheet. The rise of markerless motion capture and computer vision has promised to automate this process, but the crucial question has always been whether the automated systems see the same risks that a trained human analyst does.

To answer that question with methodological rigor, the Slovak team designed an experiment in which both platforms received exactly the same input: a single video recording of the door-removal task, broken into its constituent operations. Because the visual evidence was identical, any divergence in the resulting risk scores could be attributed not to differences in what was observed but to differences in how the two systems interpreted and scored the observed postures. The LEA application uses automated AI skeleton tracking to estimate joint angles and posture variables directly from the video, effectively letting the algorithm identify the position of the shoulders, elbows, wrists, neck and trunk across every frame. Timer Pro Professional, by contrast, follows the classical workflow of professional ergonomics software: the analyst watches the footage and manually selects the posture parameters that feed into the RULA scoring tables, a process that is transparent and deeply integrated with broader industrial engineering tools but that depends on human judgment at every step.

A key methodological contribution of the study lies in the conversion scoring framework the authors developed to make the two platforms comparable. Because LEA and Timer Pro do not natively produce identical output categories, the researchers built a scoring bridge based on the RULA method that allowed the results of both platforms to be expressed on a common scale for each evaluated operation. With that framework in place, they could measure agreement and divergence operation by operation, rather than only in aggregate. This is precisely the kind of quantitative comparison that has been missing from a literature rich in single-tool validation studies but sparse in controlled multi-platform experiments, and it reflects a broader research agenda, echoed in recent systematic reviews of digital ergonomics, that calls for systematic comparison of assessment methods rather than isolated demonstrations of new technology.

The results were unambiguous in their pattern, if not uniform in their magnitude. In four of the five evaluated operations, the two platforms produced systematically different load scores, and in every one of those divergent cases it was the AI-driven LEA platform that assigned the higher value. The largest gap appeared in the door-placement task, where LEA returned a score of 4 while Timer Pro returned a score of 2. Closer inspection of the footage revealed why: the automated algorithm had captured a transient peak of neck flexion, a brief but sharp bend of the neck at a critical moment in the movement, which the manual analysis had not flagged as a defining posture parameter. In a task that lasts only seconds, such transient extremes can be invisible to an analyst watching at normal speed, yet they may represent exactly the kind of biomechanical spike that accumulates into injury risk over thousands of repetitions.

The single operation in which the two platforms achieved absolute agreement, a score of 2 on both systems, was the mechanically simple right-foot rotation task. This contrast is telling. Where the posture is static, simple and easily described by a handful of discrete joint angles, human parameterization and AI tracking converge. Where the movement is dynamic, variable and punctuated by momentary extremes, the automated approach demonstrates markedly higher sensitivity to postural variability across time. In other words, the difference between the platforms is not primarily one of accuracy in the narrow sense of matching ground truth joint angles, which earlier validation studies of markerless systems such as Kinect-based and OpenPose-based assessments have examined extensively. Rather, it is a difference in sampling philosophy: an algorithm that scores every frame will inevitably catch fleeting extremes that a human analyst, by necessity, averages away or overlooks.

That higher sensitivity cuts both ways, and the authors are careful not to portray the automated platform as simply superior. The manual approach embodied by Timer Pro Professional offers broader multi-method support, meaning it can implement not just RULA but also other established ergonomic assessment instruments, and it integrates with the wider toolkit of industrial engineering, from time studies to process documentation. It is also auditable in a formal sense: every parameter selected by the analyst is explicit and defensible, which matters in regulatory and compensation contexts. The AI approach, meanwhile, is fast, inexpensive, and accessible from a smartphone, but its scoring depends on the quality of skeleton tracking under real workshop lighting, occlusion and clothing conditions, and it currently functions within a narrower methodological envelope. Recent scoping reviews on the automation of musculoskeletal ergonomic assessments have highlighted exactly this trade-off between the speed and coverage of automated systems and the depth and flexibility of expert-driven analysis.

It is against this backdrop that the study arrives at its most practical contribution: a proposed two-tier strategy aimed specifically at small and medium-sized enterprises, which typically lack dedicated ergonomics departments or expensive motion capture infrastructure. In the proposed model, the LEA application serves as a rapid screening filter. Because it is low-cost and requires only a video recording, it can be applied broadly across many tasks and many workplaces, flagging operations whose automated load scores exceed a threshold of concern. Only those operations identified as critical in the screening pass would then be subjected to a second-tier, detailed and formal audit using Timer Pro Professional, where an expert can apply the full range of assessment methods and produce documentation suitable for formal intervention. The authors position this as an optimal and cost-effective compromise: the sensitivity of AI screening ensures that transient risk peaks are not missed, while the rigor of expert-led second-tier analysis ensures that interventions are grounded in defensible, multi-method evidence.

The implications reach beyond the repair shop where the experiment took place. If transient postural extremes are systematically underweighted by conventional manual analysis, as this study suggests, then a substantial body of historical ergonomic assessments built on observation may have underestimated the peak loads in dynamic industrial tasks, from assembly lines to construction sites where wearable sensors and computer vision systems are increasingly deployed. The findings also give quantitative weight to the growing argument, visible across recent systematic reviews of digital ergonomics, that the field should treat automation and expert judgment not as competitors but as complementary layers of a hierarchical assessment pipeline. The Slovak study, supported by the Cultural and Educational Grant Agency KEGA 014TUKE-4/2024 and the Scientific Grant Agency VEGA 1/0302/25 of the Slovak Ministry of Education, Research, Development and Youth, demonstrates concretely what such a pipeline would look like on the floor of a working enterprise: a phone camera, an algorithm that never blinks, and a human expert deployed precisely where the algorithm says the risk lives.

Subject of Research: Multi-platform digital ergonomics assessment of dynamic physical workload using AI skeleton tracking versus manual RULA-based software in a real vehicle maintenance workplace.

Article Title: Dynamic in Situ Assessment of Work Load: Multi-platform Integration in Digital Ergonomics

Article References: Dynamic in Situ Assessment of Work Load: Multi-platform Integration in Digital Ergonomics. (n.d.). https://doi.org/10.1007/s11036-026-02554-0

Image Credits: AI Generated

DOI: 10.1007/s11036-026-02554-0

Keywords: digital ergonomics, RULA, LEA application, Timer Pro, AI skeleton tracking, computer vision, musculoskeletal disorders, workload assessment, posture analysis, automotive repair, SME ergonomics, motion capture