Key Takeaways
Real-world operational requirements made edge-based intelligence a necessity, not a technology preference.
Physical AI systems must account for real-world constraints such as privacy, connectivity, infrastructure, and environmental conditions.
The same architectural pattern behind this transit case study — sensing, inference, and decision-making at the edge — is increasingly relevant to software-defined vehicles.
Where Intelligence Lives
Software-defined vehicles are often discussed as platforms for new features, services, and digital experiences. However, one of the biggest opportunities in this space has nothing to do with features at all.
It has everything to do with where intelligence operates.
A transportation deployment designed to address safety and enforcement challenges provides a compelling answer. While the use case was transit, the architectural lessons extend far beyond public transportation.
Why Intelligence Had to Move Onto the Vehicle
A fleet safety solution wanted to address two problems: bus lane violations slowing public transportation and school bus stop-arm violations creating safety risks for children. The organization partnered with us to solve both. We needed a scalable way to identify incidents, document them accurately, and support law enforcement. Privacy requirements and regulatory obligations added another layer of complexity.
Transmitting large volumes of video for centralized processing created concerns around latency, bandwidth, cost, and responsiveness. More importantly, the system needed to generate evidence-quality results in real time. It wasn't enough to identify an event minutes later. The outcome needed to be produced while the event was occurring.
More importantly, the problem couldn't be solved effectively if intelligence lived somewhere else. Detecting a violation, validating evidence, protecting privacy, and supporting enforcement all depend on decisions happening directly on the vehicle. With all this in mind, our team quickly determined that a cloud-centric approach wasn’t going to work.
In the end, the team concluded that the best choice was to deploy edge-based computer vision running on NVIDIA Jetson devices. The system processes video directly on the vehicle, identifies violations in real time, generates supporting evidence packages, and applies privacy protections before data ever leaves the vehicle.
What Physical AI Looks Like in Practice
This deployment provides a useful example of what many now describe as physical AI.
Cameras observe the environment. AI models process events in real time. The system a) determines whether a violation has occurred, b) generates evidence suitable for law enforcement, and c) applies privacy protections before information is shared or stored. This is intelligence embedded within a physical system and operating under real-world constraints.
In physical AI systems, real-world constraints are just as important as the intelligence itself.
During deployment, the team encountered challenges that had nothing to do with model performance. In winter, temperature changes from opening and closing bus doors caused condensation to form on camera lenses, reducing visibility. The physical system itself had to be modified to prevent condensation and maintain reliability year-round.
It’s a problem that would never appear in a training dataset or lab environment. That's what makes Physical AI different. It doesn’t operate in a controlled environment, so the physical environment becomes part of the architecture.
Why OEMs Should Pay Attention
Viewed through an automotive lens, this isn’t a transit story. It’s an architecture story that gives us a preview of how future vehicle intelligence may be deployed.
Much of the first wave of SDV investment focused on how software could introduce new features and services. This deployment points to a separate category of value: vehicles that can sense, interpret, and act within the physical world. Here, value wasn't created through a subscription, a mobile application, or an infotainment experience. It was created when the vehicle could detect an event, determine whether it mattered, generate evidence, and produce an outcome in real time.
This deployment makes the concept of the vehicle as a real-time intelligent node tangible. Historically, many vehicles acted primarily as sources of data. Information was collected, transmitted elsewhere, analyzed, and acted upon later. Here, the vehicle actively participates in the process.
It senses. It interprets. It produces an outcome.
The vehicle is no longer simply generating information for later analysis. It's participating in decision-making at the point where data is created.
That architectural pattern is becoming increasingly relevant as OEMs and suppliers are increasingly pursuing:
- Distributed intelligence
- Edge computing
- Real-time inference
- Privacy-by-design architectures
- Local decision-making
- Outcome-driven data products
"If you take the labels off, this isn't a transit story — it's an automotive architecture story: sensors and compute at the edge, real-time inference, privacy controls, evidence-quality data products. That's where modern SDV programs are going."
The deployment's results support that perspective. Bus lane enforcement achieved 91% accuracy across more than 5,000 buses, while stop-arm enforcement reached 95% accuracy using multi-camera configurations. The architecture also reduced latency and minimized cloud dependency and infrastructure costs.
This important shift — from software functionality to operational intelligence — is where many SDV programs are heading.
The Future of SDVs
As vehicles generate larger volumes of sensor data and support more sophisticated capabilities, waiting for a centralized system to receive, process, and return every decision will become less practical for many use cases. Intelligence is moving closer to where data is created because that's where operational outcomes are produced.
Increasingly valuable SDV capabilities will be those that cannot tolerate delay: collision avoidance systems, driver monitoring, predictive maintenance, fleet safety applications, and other scenarios where decisions must be made in the moment rather than after a trip to the cloud.
Some problems can only be solved when intelligence is embedded directly within the vehicle — and that is where Physical AI and SDV architectures converge.
This case study was recently featured in the IDC Digital Engineering and Operational Technology Services Case Studies Showcase Report, Part 9: Physical AI Services (Doc #US54579026, July 2026), which highlighted the solution as an example of Physical AI in production.
Explore our Automotive & Industrials expertise.