Why Human-in-the-Loop Quality and Simulation-Ready Data Assets Are Non-Negotiable for Safety-Critical AI

• Annotation automation fails in safety-critical edge cases where human judgment is the only reliable signal
• While autonomous vehicle programs have matured through standardized sensor configurations and continuous collection infrastructure, robotics programs face a substantially larger annotated-data deficit driven by heterogeneous sensor stacks, episodic collection, and the absence of universally accepted annotation benchmarks
• Production-grade annotation operations depend on workforce discipline that maintains consistency across thousands of annotators and millions of sensor frames
• For enterprise teams building autonomous vehicle and robotics programs, AI data partner evaluation should cover compliance certifications, including ISO 27001, SOC 2 and TISAX

VANCOUVER, British Columbia, April 16, 2026 (GLOBE NEWSWIRE) -- Physical AI programs that fail in production almost always trace the failure back to the data layer. As of April 2026, the data picture for physical AI diverges sharply by program type. Autonomous vehicle programs have matured significantly, with standardized sensor configurations, continuous collection infrastructure, and established annotation standards producing billions of labeled frames across leading programs. Robotics programs face a fundamentally different situation: heterogeneous sensor stacks, episodic data collection, and the absence of universally accepted annotation benchmarks have left the field substantially behind, even as demand accelerates. The consequences of annotation failure are categorically different from a consumer AI application simply getting something wrong. A misclassified object in a lidar point cloud represents a potential safety failure. The annotation operations that excel in production share six qualities that are easy to overlook in a pilot. TELUS Digital, a global leader in AI data solutions for vehicle and robotics programs, has worked through all six of them. 

Steve Nemzer, Senior Director, Artificial Intelligence Research & Innovation at TELUS Digital, says, “Pilots can be gold-plated with manual processes and hand-picked people—they prove feasibility. Production-grade annotation operations work across diverse teams, at scale, with the discipline to enforce consistency. They prove repeatability. The gap between pilots and production is the ability to manage at-scale workforces without sacrificing quality.”

KEY FACTS:

• TELUS Digital's AI Community includes more than 1 million trained data annotators and linguists across six continents
• TELUS Digital delivers more than 2 billion labels annually in 500 or more annotation languages
• Safety-critical compliance requirements for AI data partners include ISO 27001, TISAX, ISO 31700-1, HITRUST, SOC 2 and GDPR/CCPA
• TELUS Digital's Ground Truth Studio platform supports camera-lidar fusion, 3D point cloud segmentation and lane detection in 2D and 3D scenes

What Makes Safety-Critical Annotation Different

Enterprise teams building autonomous vehicles and robotics are facing challenges that consumer AI development doesn't impose. Annotation quality in physical AI exemplifies this challenge—it is not a mere background variable. Incorrect actions in a physical environment have physical repercussions. A pedestrian identified in a lidar point cloud must correspond precisely to the same pedestrian in the camera frame and the radar return. Cross-modal consistency failures produce perception models that generate conflicting readings of the same scene. In an autonomous vehicle, that conflict is a safety risk. In a robotics context, it produces a failure to act or an incorrect action.

The following reflect production-ready annotation best practices at a safety-critical scale:

1. Human Judgment at the Boundary of Automation

Automated annotation handles high-volume, repetitive labeling well, but it struggles with ambiguous or rare edge cases. In real-world scenarios, ambiguity is high and the cost of error is unacceptable.

“Annotated automation hits a wall in those safety-critical edge cases where ambiguity is high. For example, interpreting the gesture of a crossing guard is far trickier than identifying a yield sign. Annotation processes at scale don't try to automate away human judgment. Automated systems flag high-uncertainty cases (using confidence thresholds, disagreement signals, etc.) and expert human-in-the-loop annotators resolve them with structured decision frameworks,” Nemzer says.

Production annotation pipelines for physical AI are designed to keep moving. When automated systems encounter high-uncertainty cases they can't resolve reliably, those cases are routed to human experts. The pipeline stays efficient by letting automation handle the straightforward issues while concentrating human effort exactly where judgment is needed.

2. Cross-Modal Consistency Across Lidar, Radar and Camera

Annotation platforms that exclusively handle one or two sensor types or treat fusion as a secondary step generate misaligned training data that permeates the dataset. For L4+ autonomous vehicle programs, where the perception stack must perform reliably at highway speeds across all weather conditions and geographies, cross-modal inconsistency is a direct risk to the program. 

One of the most common sources of misalignment is temporal drift. Even a 50-millisecond gap between sensor captures means a pedestrian detected at frame N in the camera feed may appear at frame N+2 in the lidar return, creating a ghost object that the perception model has no reliable way to resolve. At highway speeds, that gap translates directly into a labeling error that propagates through training. Production-grade annotation operations address this through automated temporal alignment checks that ensure every object labeled in camera data has a verified corresponding label in lidar and radar. For enterprise AV teams, this is one of the failure modes that experienced annotation partners know to look for and that general-purpose labeling platforms are not designed to catch.

Training autonomous vehicles and robots requires labeling data from multiple sensors, with every object labeled consistently across all of them simultaneously. TELUS Digital's Ground Truth Studio was built for this level of complexity. It supports camera-lidar fusion, 3D point cloud segmentation, compatibility across solid-state and flash lidar sensors and automated object interpolation for video annotation at scale.

3. Simulation Pipeline Readiness for World Model Development

Synthetic data generated in environments like NVIDIA ISAAC-Sim is effective for training embodied AI systems. However, models trained purely in simulation encounter a fundamental physics gap in real-world deployment. Many simulation environments use simplified approximations such as point contacts, linearized friction models and stable surface assumptions to maintain computational efficiency. Real-world contact is inherently nonlinear: materials deform under load, friction varies with velocity and contact states shift unpredictably between sticking and slipping. Debris, surface irregularities and stochastic environmental dynamics compound this further. As a result, grasps that succeed in simulation become unstable in deployment, and motion that appears reliable under controlled conditions breaks down against real physical variability. No simulator currently replicates these behaviors at the fidelity required for production-ready physical AI.

"The balance to strike is to use synthetic data to fill specific data gaps while anchoring training on real-world data that grounds the model in the long tail of real-world variability. Synthetic data can't teach models about the sensor artifacts or adversarial conditions they'll encounter in production,” Nemzer says.

Simulation-ready data pipelines need more than synthetic generation. They need human-in-the-loop annotation to capture what simulation misses and quality systems to keep both data types consistent.

4. Production-Scale Workforce with Domain Expertise

The distinction between pilot-scale and production-scale annotation isn't really a technology problem. Pilots can be managed with manual oversight and hand-selected annotators. Production programs require active learning systems, consensus annotation workflows, multi-stage quality review and infrastructure to enforce annotation guidelines consistently across thousands of annotators working on millions of sensor frames.

For physical AI programs, domain expertise in annotators directly improves data quality. A team that understands the underlying technology (sensors, kinematics, safety requirements and risks) produces better training data because they understand why each label matters.

5. Data Lineage and Traceability from Raw Sensor Input to Labeled Output

Production-grade data operations for safety-critical AI programs demand full traceability. 

"Data lineage isn't a nice-to-have for safety-critical AI. You need to be able to quickly answer questions like what exact data trained this model, what quality standards did it meet and why did it fail in this specific case, without extensive manual investigation. If you're having to dig through logs, you're not ready for production safety-critical work," Nemzer says.

Data lineage and version control in the annotation pipeline include:

• Ingestion records
• Preprocessing logs
• Annotation guideline versioning
• Quality assurance records
• Delivery documentation

6. Compliance Certifications Aligned to Program Requirements

Safety-critical AI programs in automotive, robotics and industrial applications carry compliance requirements that generic annotation vendors may not be able to meet. Core certifications for AI data services partners in these programs include:

• ISO 27001 for information security management
• TISAX for automotive-specific data handling 
• ISO 31700-1 for privacy by design
• HITRUST for healthcare-adjacent applications
• SOC 2 Type 2 for service organization controls
• GDPR and CCPA/CPRA for data privacy compliance.

What the Criteria Tell Procurement Teams

Procurement teams must address all six considerations simultaneously, as gaps in any one area will compound during model training. While autonomous vehicle programs have matured in annotated dataset scale, the robotics data gap remains substantial and will close as collection operations and annotation standards mature. Building data operations on quality systems designed to handle that scale from the start will help programs reach production sooner. 

FAQ:

Q: What should we look for in human-in-the-loop annotation services for a multi-modal AI system?

A: For multi-sensor programs, native cross-modal annotation support across lidar, radar and camera-lidar fusion is a baseline requirement. Domain expertise in the relevant sensor modalities determines whether the training data holds up at deployment.

Q: What does edge case data collection for safety-critical AI really require?

A: Real-world collection captures sensor artifacts and long-tail variability that simulation cannot replicate. Synthetic data covers scenarios too infrequent to collect at scale, including construction zones and emergency vehicle interactions. Therefore, both data are necessary. Additionally, edge case datasets need the same quality standards and audit trail requirements as primary training data.

Q: Which annotation capabilities matter most for complex robotics applications?

A: Native support for 3D bounding boxes, semantic segmentation, panoptic segmentation and temporal sequence labeling across fused sensor data is the starting point. Force and torque sensor inputs and state-action-behavior data used in visual-language-action model training are also worth verifying before selecting a partner.

Q: What separates a production-ready annotation operation from one that breaks at scale?

A: Pilots run on manual oversight and hand-selected teams, whereas production programs need active learning systems and multi-stage quality review. When a model fails, the team needs to be able to trace that case back to the training data quickly, without reconstructing the audit trail from scratch.

Q: What compliance certifications should an AI data partner hold for safety-critical applications?

A: ISO 27001, TISAX, SOC 2 Type 2, GDPR, and CCPA are the baseline certifications worth reviewing before a partner is selected. Programs operating under EU AI Act governance for high-risk systems should confirm if partners maintain documented audit trails and data provenance tracking as active operational requirements.

About TELUS Digital
TELUS Digital, a wholly-owned subsidiary of TELUS Corporation (TSX: T, NYSE: TU), crafts unique and enduring experiences for customers and employees and creates future-focused digital transformations that deliver value for our clients. We are the brand behind the brands. Our global team members are both passionate ambassadors of our clients’ products and services and technology experts resolute in our pursuit to elevate their end customer journeys, solve business challenges, mitigate risks and drive continuous innovation. Our portfolio of end-to-end, integrated capabilities include customer experience management, digital solutions, such as cloud solutions, AI-fueled automation, front-end digital design and consulting services, AI & data solutions, including computer vision and trust, safety and security services. Fuel iX^TM is TELUS Digital’s proprietary platform and suite of products for clients to manage, monitor and maintain generative AI across the enterprise, offering both standardized AI capabilities and custom application development tools for creating tailored enterprise solutions.

Powered by purpose, TELUS Digital leverages technology, human ingenuity and compassion to serve customers and create inclusive, thriving communities in the regions where we operate around the world. Guided by our Humanity-in-the-Loop principles, we take a responsible approach to the transformational technologies we develop and deploy by proactively considering and addressing the broader impacts of our work. Learn more at: telusdigital.com.

CONTACT: Sarah Evans
Partner, Head of PR, Zen Media
sarah@zenmedia.com