How BMW Group is driving major efficiencies with AI and virtual teardowns
Overview
Virtual Teardown Assistant combines real-time 3D data with LLMs to automate the benchmarking process.
Use case
AI-driven vehicle component analysis using accurate 3D image renders.
Solutions
Unity Asset Manager, Unity Pipeline Automation, Unity Asset Transformer SDK
Customer
BMW Group
In the evolving landscape of automotive engineering, understanding intricate product details and optimizing every component is more crucial than ever. Traditionally, this involved physical inspections, an approach that is resource-intensive, time consuming, and hard to scale
But what if you could tear down a car’s component parts without ever touching a wrench?
This is the question that drove Dr. Savo Asanin, technical project lead for AI-driven activities within BMW Group’s Development Strategy department, to look for a better way. The mission is clear: compare design concepts through rigorous technical analysis of component parts.
By collaborating with Unity, Savo and his team have moved beyond physical constraints. They created the Virtual Teardown Assistant, a new tool that combines real-time 3D data with Large Language Models (LLMs) to automate the design concept comparison and benchmarking process. In this case study we hear from Savo, and Unity Senior Software Development Consultant, Daniel Bierwith how it was created.
From physical limitations to digital scalability
As Savo explained, for years component analysis at BMW Group relied heavily on manual inspection. Engineers examined parts physically or searched databases for the information they needed. The approach worked, but it was resource-intensive, time-consuming, and difficult to scale.
The challenge wasn't a lack of data; BMW has massive repositories of CAD data. The challenge was accessibility and interpretation. Comparing the geometry of different components across vehicle generations required significant labor resources.
This is where the foundation of BMW Group’s 3D Mine comes in. As detailed in our previous story on BMW’s asset management solution, 3D Mine serves as a centralized solution that helps democratize 3D data access across the organization.
At a basic level, the Virtual Teardown Assistant builds on that foundation and adds a layer of intelligent automation.
The objective was to create a tool that could compare geometries of car components to find potential improvements automatically - and to achieve this with high scalability, letting the machine do the heavy lifting of analyzing components.
The engine room: Unity Asset Manager and Pipeline Automation
To build the Virtual Teardown Assistant, you need more than just a 3D model; you need a system that can understand what it is looking at.
The workflow starts with complex CAD data, often an entire vehicle model. Using Unity Asset Manager and Pipeline Automation, the system breaks the vehicle down into its constituent systems such as a seat or a braking system.
As Savo explained, the workflow is to take a CAD model, explode it into its parts, color-code them, and essentially 'photograph' them against a white background in the Editor. All of which takes place automatically.
This is where Unity’s enterprise solutions shine:
- 3D Mine, powered by Unity Asset Manager: Acts as the single source of truth, retrieving selected 3D models and associated metadata files.
- Pipeline Automation: Handles the heavy lifting. It isolates parts based on data, explodes the view so every screw and bracket is visible, and renders high-quality images for analysis.
The generated images, along with a Bill of Materials (BOM) containing crucial metadata like part numbers and dimensions, are fed into an LLM. The AI "looks" at the exploded view, reads the data, and provides an engineering analysis, identifying opportunities for weight reduction, cost savings, or functional improvements at scale.
Surprising insights and significant cost savings to date
When the team first launched the Proof of Concept (POC), they weren't sure how the AI would handle visual data.
Savo’s team were unsure what to expect, but they were pleasantly surprised at how well the model worked with limited geometric information. It produced valuable insights and understood the geometry. The team is already seeing positive results on prioritized systems.
The implications of this efficiency are significant. By automating the pipeline for the whole vehicle across different systems, BMW is increasing efficiency and enabling faster engineering analysis.
Rolling the Virtual Teardown Assistant out across the organization allows teams to compare concepts, designs, and supplier solutions more quickly and with greater consistency. This supports faster benchmarking and helps engineers identify the most promising options earlier in development.
Collaborating to innovate
The success of the Virtual Teardown Assistant highlights the deepening enterprise relationship between BMW Group and Unity. It’s not just about buying software; it’s about co-developing solutions.
Savo’s team collaborated closely with Unity to build this solution, with access to technical expertise directly from Unity’s customer-facing teams. This resulted in a platform that is easy to onboard, and the performance in visualizing parts in high quality is exactly what the team hoped for.
For the end-user at BMW, the engineer making critical design decisions, the complexity of the backend is invisible. They don't see the Unity solutions such as Asset Transformer SDK converting files or the Pipeline Automation scripts running in the background. The feedback from the POC has been overwhelmingly positive, with teams eager to use the tool on a wide range of vehicle components.
Example showing labeled systems in the context of a car model. Courtesy of BMW Group.
Technical deep dive: How the Virtual Teardown Assistant was built
Curious about how this transformative tool was created? In the following section, Daniel Bierwith, Senior Software Development Consultant at Unity, breaks down the architecture, technologies, and engineering approach that brought the Virtual Teardown Assistant to life.
The essential components
The Virtual Teardown Assistant is powered by Unity Asset Manager, a scalable platform that streamlines the creation of detailed explosion renders. By connecting directly to the LLM pipeline through automated APIs, the system transfers output data effortlessly.
At the heart of this solution is Unity Pipeline Automation, a low-code interface that links systems, transforms data, and updates 3D content at scale. Pipelines run in virtual machines on a flexible cluster and can be triggered by events, schedules, or on demand. For more information, see the Appendix below.
The process is accelerated by Unity Asset Transformer SDK, which supports 70+ 3D and CAD file formats. It handles ingestion, optimization, and standardization, making it easy to integrate assets into any real-time 3D workflow.
A centralized Digital Asset Management hub (in this case, BMW Group’s 3D Mine) keeps everything organized and accessible. Teams can retrieve, manage, and collaborate on complex real-time 3D assets in one place, speeding up production and enabling the creation of scalable, high-quality digital experiences.
Input data is uploaded to the Unity Asset Manager to create a central shared repository for collaboration on the Virtual Teardown Pipeline. Courtesy of BMW Group.
Virtual Teardown Assistant pipeline steps
The Virtual Teardown Assistant pipeline streamlines the process of analyzing complex 3D engineering models by combining automated data processing, visualization, and AI-driven insights.
Pipeline overview
The workflow begins when the Model Load Action retrieves the selected 3D engineering model in JT format from the Unity Asset Manager/3D Mine. Simultaneously, a system filter file is loaded to isolate the specific subsystem within the full product model. Both actions are configured through the graphical pipeline interface.
In a typical use case, the pipeline processes JT files exceeding multiple GB that contain complete engineering models with thousands of components and millions of polygons.
Pipeline steps with asset retrieval and Asset Transformer execution to isolate and render the exploded views.
Data processing and transformation
Once the model is loaded from Asset Manager, the pipeline executes the main processing tasks through a custom Python script running on a virtual machine equipped with the Asset Transformer SDK. The script loads the complete JT model. Asset Transformer SDK is used to apply a filtering mechanism to the occurrences of the data set to isolate the relevant system parts using the filter table imported from the CSV file in the previous pipeline step.
Next, the SDK algorithm is used to extract and adjust material properties to assign unique colors to each component for clear visual differentiation. This is the labeled occurrence representation that will be used by the LLM alongside with the BOM and geometric properties.
Example showing labeled front axle system in the context of a car model. Courtesy of BMW Group.
The isolated components are exploded using a custom algorithm that accounts for both hierarchical and geometric relationships. For the visual analysis by the LLM, it is important to maintain proper separation of parts and ensure that they are clearly visible to the capture camera.
The Pipeline interface allows the application to apply custom settings as explosion and render recipes to dynamically modify the pipeline settings based on product and system characteristics.
The Asset Transformer SDK provides cloud rendering capabilities that are used to render the labeled exploded views of the system as part of the processing pipeline. The resulting visual data serve as the foundation for both human review and AI analysis.
Example explosion render of the front axle system. The left image shows the isolated system, the right image shows the generated exploded view. Courtesy of BMW Group.
Data export and AI analysis
Additionally, a custom scripting function exports the BOM as a CSV file containing metadata for all system components and calculates the geometric characteristics of each occurrence using the Asset Transformer SDK functions. The BOM is stored alongside the rendered exploded views in the 3D Mine.
Example explosion renders for the front bumper system. The left image shows the isolated system, the right image shows the generated exploded view. Courtesy of BMW Group.
The AI component takes over. The Pipeline Automation interface sends HTTP requests to a large language model (LLM), providing the exploded views, BOM, and system description:
- First, a pipeline step creates the payload for the LLM request by encoding the rendered images and the BOM in the request body, along with the system description, LLM instructions, and the prompt.
- The LLM performs a comprehensive analysis of the input data and returns text files containing key architectural insights, materials and joining strategies, manufacturability assessments, first-order performance KPI estimations, assumptions with confidence ratings, and design improvement suggestions.
Virtual Teardown Assistant pipeline steps to set up connections via API to the LLM pipeline, facilitating the transfer of pipeline output data without manual intervention.
Results and impact
The final pipeline step uploads the AI-generated benchmark report and labeled exploded views back into the Asset Manager, providing a centralized reference for teams. This enables early benchmarking before any physical teardown occurs.
The Virtual Teardown Assistant pipeline automates complex CAD processing and visualization tasks while integrating AI-driven insights, creating a powerful tool for early-stage product analysis. Its combination of scalable pipeline automation, real-time rendering, and advanced LLM evaluation significantly reduces the time and cost required to obtain actionable engineering benchmarks.
The road ahead
The Virtual Teardown Assistant is currently deployed and delivering consistent renders for high-priority systems. But Savo and his team are just getting started.
Over the next two years, the plan is to optimize the system for scale and connect the tool directly to BMW’s internal systems, such as their CAD repository. This will allow for even more granular analysis taking pictures from different angles and focusing on single components to improve quality. Savo’s team learned that combining vision algorithms with large language models is the key to scalability.
For other companies interested in adopting similar solutions, the advice is clear. Pick a real problem, build a fast POC, and let the results speak. Strategy matters, but nothing convinces an organization like seeing it work. Once the impact is visible, scaling becomes a pull from the business, not a push from IT.
By leveraging Unity’s ecosystem, BMW Group isn't just visualizing cars; they are teaching machines to understand them, driving a smarter, faster, and data-focused future for automotive engineering.
