Ergonomic Assessment Software: How AI Video Tools Work, What They Get Right and Where They Fall Short

What is AI-based ergonomic assessment software?

AI-based ergonomic assessment software estimates a worker's postural risk from video or images instead of having an assessor measure joint angles by eye. Most tools in this category do three things:

1. Detect the body. A pose estimation model finds joints such as shoulders, elbows, hips and knees in each frame.
2. Calculate angles. The software converts those joint positions into trunk flexion, upper arm elevation, neck angle and so on.
3. Score a method. The angles are mapped onto an established observational method, most often REBA or RULA, to give a risk score.

The appeal of this kind of ergonomic assessment software is speed and consistency. A manual REBA on a multi-step task can take a long time and two assessors often disagree. Software can process the video in minutes and gives the same answer every time it sees the same footage.

ViveLab's AI ergonomic assessment works this way: you upload a phone video or photo, pose detection runs in the browser on your device, and you get an estimated REBA or RULA score with a report. Like every tool in this category, it produces an estimate that a trained assessor should review.

What kinds of ergonomics software are there?

"Ergonomics software" covers several different approaches. They are not interchangeable.

Type	How it captures posture	Strengths	Trade-offs
Video and pose estimation	Ordinary RGB camera or phone, AI detects joints	Cheap, fast, no equipment on the worker	2D view, occlusion, camera-angle sensitivity
Wearable inertial sensors	Small sensors strapped to body segments	Measures angles over a full shift, no line of sight needed	Hardware cost, setup time, workers must wear them
Depth and multi-camera systems	Depth sensors or several synchronized cameras	Better 3D estimates than a single camera	Fixed setup, harder to move around a site
Marker-based motion capture	Reflective markers and lab cameras	The research reference standard	Lab-bound, expensive, not practical on most floors
Checklist and survey software	Assessor answers structured questions	Covers force, load and duration, not just posture	Still relies on human judgment

A 2025 systematic review in Sensors compared marker-based and markerless camera systems for industrial ergonomic risk analysis. It found marker-based systems had the highest precision and reliability but were limited by cost and lab constraints, while markerless systems showed moderate-to-high accuracy, good reliability and greater practicality in the field (Scataglini et al., 2025).

How does pose estimation work in practice?

Pose estimation models are trained on large sets of labeled images to predict where body landmarks are. Google's MediaPipe Pose Landmarker, for example, tracks 33 body landmarks and outputs them both in image coordinates and as estimated 3D world coordinates. Google describes the model bundle as optimized for on-device, real-time use, and provides a JavaScript version that runs in a web browser (Google AI Edge).

From those landmarks, the software computes segment angles frame by frame. It then picks the frames that represent the worst or most frequent postures, applies the scoring tables of the chosen method, and reports a score.

The scoring tables are the published methods. The AI part is the angle measurement and frame selection. If you want to check the arithmetic, you can enter the same angles into a manual REBA calculator and compare.

How accurate is video-based ergonomic assessment software?

The honest answer: it depends on the body part, the camera view and the task. Here is what peer-reviewed studies report.

Consistency is very high

In a 2025 PLoS One study, three human raters (novice, intermediate and expert) and a commercial deep-learning video tool scored REBA on 21 veterinary task videos. The AI tool's intra-rater reliability for the final REBA score was 1.00, compared with 0.89 for the expert, 0.66 for the intermediate rater and 0.51 for the novice. Agreement between the three humans was low, with an inter-rater ICC of 0.36 (Balogh et al., 2025).

That consistency is real value. Human REBA scores vary a lot between people.

Agreement with experts varies by body segment

In the same study, agreement between the AI tool and the expert was highest for the trunk, leg and upper arm, and lowest for the neck, wrist and lower arm. When the expert scored still frames chosen by the software, weighted kappa for the trunk was 0.63 and for the legs 0.57, while the wrist was close to zero (Balogh et al., 2025).

The authors note that they did not assess validity, so the study cannot say whether the AI or the humans were closer to the true angles. They also suggest the side and oblique camera views made wrist posture hard to judge.

Angle accuracy against motion capture is promising

The Sensors review reported a mean joint-angle error of 2.31 degrees (plus or minus 4.00) for markerless camera systems, good reliability (ICC above 0.80), and several studies reporting strong validity for RULA and REBA prediction, with accuracy up to 89 percent and kappa of 0.71. It also flagged that many studies had fewer than 10 participants and that reliability was underreported (Scataglini et al., 2025).

An earlier study compared an OpenPose-based system with a reference motion capture system across 10 participants and 12 tasks, with and without body occlusions and from frontal and non-frontal views. OpenPose performed well under all conditions, while a Kinect depth-sensor system did significantly worse with occlusion or non-frontal tracking (Kim et al., 2021).

For a deeper look at the numbers, read our post on pose estimation accuracy in ergonomics.

What can video-based software not see?

These limits apply to every single-camera tool, not one product.

• 2D projection. A single camera flattens 3D movement. A trunk bent toward the camera looks less flexed than it is. Some models estimate depth, but it is still an estimate.
• Occlusion. Arms behind the body, a worker behind a bench, or another person in frame can hide joints. The software has to guess.
• Camera angle. Side views capture flexion well but miss twisting and side bending. Front views do the opposite.
• Wrist and hand detail. Wrist deviation and grip are small movements that full-body models track poorly, which matches the low wrist agreement in the study above.
• Load weight and force. Video cannot see how heavy a box is. REBA and RULA both add points for load or force, and the NIOSH lifting equation depends directly on weight. OSHA notes that weight alone does not determine lifting risk, but it is a required input (OSHA, 2013).
• Coupling, activity and duration. Grip quality, static holding and repetition across a shift are partly visible at best.

Good software asks you to enter the missing inputs instead of guessing them. If a tool reports a full REBA score without asking about load, ask how it handled that.

How should you evaluate ergonomic assessment software?

Use this checklist when comparing tools:

1. Methods. Which methods does it score (REBA, RULA, NIOSH, others)? Does it follow the published scoring tables exactly?
2. Manual inputs. Can you enter load, force, coupling and activity? Can you correct an angle the AI got wrong?
3. Frame selection. Can you see and change which frames were scored?
4. Transparency. Does it show measured angles per body segment, not just a final number?
5. Evidence. Is there peer-reviewed or published validation, and under what conditions (camera view, task type, sample size)?
6. Privacy. Where is video processed and stored? Can it run on the device? How long is footage kept?
7. Reporting. Does the report include the inputs, the frames and space for assessor sign-off?
8. Fit for your site. Test it on your own tasks, lighting and camera positions before deciding.

Avoid relying on any vendor accuracy percentage unless you know what it was measured against and on which tasks.

How should AI fit into your assessment workflow?

The best use of AI-based ergonomic assessment software is to widen the funnel and keep humans on the decisions:

1. Screen many jobs quickly with video to find the likely high-risk ones.
2. Review the AI output: check angles, correct obvious errors, and add load and force.
3. Confirm high scores with a trained assessor, using a full method where needed. Our ergonomic risk assessment tools guide helps you pick one.
4. Fix the job, prioritizing engineering controls.
5. Re-record after the change to show the score dropped.

ViveLab is built for this workflow. Processing happens on the device, so footage does not need to be uploaded to a server, and the report is designed to be reviewed and signed off by an assessor rather than treated as a final verdict.

Next steps

Record one task you already know well and run it through AI posture analysis, then compare the result with your own manual score using our REBA assessment guide. If you are choosing an app for regular use, see our guide to picking a posture analysis app.