Closing the Sim-to-Real Gap Enables Physical AI at Industrial Scale

Aaron Blutstein 20 July 2026  

For over two decades, simulation has played a key role in helping companies across a range of sectors to accelerate the development and deployment of robotic automation throughout their processes. Enabling users to build, simulate and refine robotic applications in a virtual world, simulation software tools, including ABB’s RobotStudio suite, have proven invaluable in reducing risk and shortening commissioning time.

While these tools have proven highly successful, one persistent obstacle has limited their progress, especially as a new generation of AI-driven robots emerges.

Traditional simulation and the ‘sim-to-real gap’

Traditional robotics simulation tools have largely focused on geometry and motion planning based on modelling kinematics, reachability, and collision detection with high accuracy. While essential, this approach overlooks the dynamic and variable nature of typical industrial environments. As more users turn to AI-enabled robots, it is becoming increasingly necessary to also factor in variations that could affect the robot’s performance in a real production situation.

While traditional virtual simulation tools model robots in perfect production conditions, the reality of many factory settings is anything but. Shifts in lighting intensity, inconsistent surface reflections, misalignment of parts, and sensors introducing distortion and noise can all impact on the ability of AI vision systems trained under ideal virtual conditions.

Controller behaviour can also cause complications. By approximating robot motion rather than executing the same controller logic used on the physical robot, traditional simulation environments can introduce timing differences, path deviations, and acceleration profiles that can affect performance.

Known as the ‘sim-to-real gap’, this gap between the virtual and real worlds can have major implications, especially for producers in sectors such as consumer electronics, automotive manufacturing, logistics, and pharmaceuticals that rely on micron-level precision and high throughput. Problems can halt production lines, increase scrap rates, and delay commissioning by weeks or even months, not to mention impairing competitiveness.

Although Physical AI promises flexibility through its ability to perceive, decide, and act autonomously, most AI-driven automation struggles to scale beyond controlled environments. The missing foundation is not more AI models, but rather the ability of industrial grade simulation to design, train, validate, and deploy robotic systems that behave in real life exactly as they do in the virtual world.

Closing this gap between simulation and reality has become one of the central challenges in scaling physical AI. One way to do this is to create a simulation environment that integrates product design, perception, and robot execution into a single, concurrent engineering workflow.

Tackling the challenges of scaling Physical AI

ABB’s RobotStudio HyperReality addresses these challenges by transforming simulation into a production-grade engineering environment where each aspect is considered simultaneously.

Through the integration of ABB’s proven RobotStudio Suite with NVIDIA’s Omniverse real-time 3D development open platform, it is now possible for individuals and teams to build and operate industrial-grade, physically accurate digital twins and virtual worlds.

Complete simulations can be created that can model geometry and motion as well as variations in materials, lighting behavior, optical properties, and sensor characteristics to an accuracy of 99 percent. This perceptual accuracy significantly reduces the mismatch between what an AI model experiences during training and what it encounters in production.

Controller fidelity as the foundation

The software incorporates ABB’s virtual controller, enabling robotic motion to be simulated with extremely high accuracy by using the same controller logic that governs the physical robot to also run the simulation.

This ensures that trajectory planning, acceleration, timing, and execution constraints behave identically in both environments. What is validated virtually becomes directly transferable to the shop floor, minimizing commissioning surprises and reducing positioning errors to sub-millimeter levels.

Synthetic data and engineered variability

In the same way that humans apply learned experience to carrying out tasks, HyperReality uses large-scale synthetic data generation to expose AI systems to thousands of realistic variations in lighting, part positions, material finishes, and edge cases. By proactively engineering variability into training, robots are conditioned to handle unusual part orientations or lighting anomalies before deployment. This aids reliable autonomy by ensuring that robots can be better prepared to meet unexpected changes, as they will have already encountered many of the situations that could cause problems on the factory floor.

Economic and operational impact

Closing the sim-to-real gap in this way reduces reliance on physical prototyping and shortens commissioning cycles, opening new opportunities for companies to reduce development costs, accelerate time-to-market, and deploy AI-powered robotics with greater confidence.

The convergence of photorealistic simulation, controller-level fidelity, and synthetic variability enables manufacturers to scale physical AI from experimental projects into dependable industrial systems.

The benefits that RobotStudio HyperReality can deliver are demonstrated by a pilot application at a major electronics manufacturer, which is using it to optimize its product assembly line. Its multiple product variants require different assembly methods, with the delicate structure of items such as buttons presenting significant challenges that can disrupt production.

To tackle this, the company is using the tool to train its assembly robots virtually with synthetic data to perfect multiple real-world production scenarios before transferring the process to the factory floor.

The improved performance this is delivering in production is helping to cut setup times, eliminate the time needed for training and testing, and speed up time-to-market.

From virtual confidence to physical autonomy

A continuous feedback loop between digital and physical systems helps sustain performance over time. AI models can be trained and validated in simulation using extensive computing capacity, then refined after deployment using real production data to keep the digital twin aligned with operating conditions.

ABB’s OmniCore controller platform further supports AI-intensive workloads by incorporating NVIDIA-based computing hardware capable of processing advanced vision and sensor data. This architecture ensures that the robustness achieved in simulation can be maintained under real-time production constraints.

Making Physical AI happen

The sim-to-real gap has long been the primary barrier preventing AI-powered robotics from scaling beyond controlled demonstrations. Combining the key elements needed for a robot to act autonomously such as photorealistic perception modelling, controller-level execution fidelity, large-scale synthetic variability, and integrated hardware support, directly addresses the root causes of this gap to make physical AI a dependable reality on the factory floor.