Expert Perspectives from Across the Physical AI Ecosystem

One of the most thought-provoking discussions at CadenceLIVE centered on a challenge that sits at the heart of modern system innovation—how do we ensure that what works perfectly in simulation performs just as reliably in the real world?

The panel on “Closing the Sim-to-Real Gap in Physical AI” brought this question into focus, exploring the rapid evolution of physical AI across robotics and autonomous systems and the very real hurdles to scaling it.

Bringing together a diverse group of industry leaders, the panel featured Suraj Gajendra, VP of Products and Solutions for the Physical AI Business Unit at Arm; Mahesh Kailasam, corporate VP of R&D at Cadence; Monica Xie, head of Partnership at DYNA Robotics; Amit Goel, head of Robotics and Edge Computing Ecosystem at NVIDIA; and Prasun Raha, VP of Systems at Rivian. Moderated by Frank Ham, fellow at Cadence, the discussion provided a comprehensive exploration of the ecosystem necessary to drive physical AI beyond controlled environments.

Highlights from the Panel Discussion

The panel, moderated by Frank Ham, explored the evolving landscape of Physical AI and the persistent challenge of closing the sim-to-real gap. Bringing together leaders across simulation, silicon, robotics, and autonomous systems, the discussion examined whether Physical AI is approaching a “ChatGPT moment,” and what still stands in the way.

Frank set the stage by reflecting on the LLM revolution as a convergence of GPU hardware, transformer-based software mechanisms, and vast amounts of text data that led to a step-function leap in capability. He contrasted this with Physical AI, where the problem extends beyond perception and language into real-world interaction—recognizing objects, manipulating them, and responding to physical dynamics. This complexity introduces the sim-to-real gap as a central challenge.

Mahesh Kailasam emphasized that Physical AI is fundamentally harder than LLMs. Even if intelligence and knowledge are available, systems must physically execute actions and respond to the environment in real time. The physical embodiment—motion, mechanisms, and interaction with dynamic conditions like terrain or materials—adds layers of complexity absent in purely digital systems. While simulation has advanced significantly, approximations in modeling—such as treating flexible objects as rigid or simplifying contact interactions—accumulate and lead to breakdowns when systems move into the real world.

Amit Goel framed the challenge through three key differences. First, Physical AI is inherently multi-modal, involving forces, friction, audio, and other continuous data types, making the data landscape far more complex than text-based systems. Second, there is a lack of structured, usable data; while the physical world is constantly being captured, converting it into forms that models can learn from remains difficult. Third, deployment requires edge computing, where intelligence must operate continuously on-device in dynamic environments. Simulation becomes essential—not only for generating scalable data but also for validation and safe testing, which cannot be done through simple feedback loops as in digital systems.

Monica Xie pointed to a gap between the AI and robotics communities. Robotics is benefiting from advances in large models, but the exchange of lessons between the two domains remains limited. AI practitioners often underestimate the complexity of hardware and control systems, while robotics teams have yet to fully adopt scaling strategies from large model development. Simulation, in this evolving context, is becoming more valuable—not necessarily as a perfect representation of reality, but as a way to efficiently test general-purpose capabilities across a wide range of scenarios.

Suraj Gagendra highlighted the importance of scale and system-level constraints. Simulation environments must be rich and extensive enough to support large-scale deployment across many devices. At the same time, compute challenges—such as energy efficiency, real-time determinism, safety, and security—add further complexity. As robotics shifts from fixed-function tasks to multi-task, adaptive systems, simulation becomes critical for preparing systems to operate in unpredictable environments.

Prasun Raha provided an automotive perspective, where simulation plays a crucial role across the lifecycle—from design to deployment. It helps resolve trade-offs such as sensor placement while maintaining product design integrity. In deployment, extensive simulation—running millions of miles of virtual driving—supports validation before releasing updates. However, simulation alone is insufficient; real-world testing remains necessary to capture edge cases and ensure correct behavior under real conditions.

A consistent theme across the panel was the need to balance speed and accuracy in simulation. Faster simulations enabled by GPU acceleration allow for scale and broader scenario testing, but without accurate physics, results cannot be trusted. High-fidelity, engineering-grade simulation provides the necessary foundation, and both speed and accuracy must evolve together to close the sim-to-real gap.

The discussion also explored why this moment is different. Advances in large models have demonstrated how to learn from unstructured data at scale. At the same time, improvements in simulation—particularly GPU-accelerated physics—now allow years of learning to be compressed into much shorter timeframes. Additionally, the ability to bring real-world data back into simulation environments is improving fidelity and reducing the gap between simulated and real experiences. These converging developments are enabling progress that was previously not possible.

Dexterity emerged as a key challenge, especially in robotics. Tasks involving deformable objects, such as fabric handling, remain difficult to model and simulate accurately. However, new approaches to data—combining large-scale, lower-cost pre-training data with smaller amounts of high-quality task-specific data—are changing how these problems are addressed. Simulation, in this context, is evolving into a scalable source of training data rather than a perfect replica of reality.

Safety was a central concern throughout the discussion. Unlike digital AI systems, where errors may be inconsequential, failures in Physical AI systems can have severe real-world consequences. As robots move out of controlled environments and into spaces shared with humans, safety must be addressed across hardware, software, and system design. Simulation plays a key role in testing rare and extreme scenarios, but it must be complemented by systems capable of reasoning, predicting, and adapting in real time.

The panel also emphasized the importance of moving from task-specific systems to general-purpose intelligence. As systems take on more complex, multi-step tasks over longer time horizons, safety and control can no longer rely solely on predefined rules. Instead, systems must develop contextual understanding and decision-making capabilities similar to human reasoning.

The discussion closed with forward-looking predictions. Mahesh anticipated “simulation-powered intelligence on the edge,” where real-time decisions are informed by embedded simulation. Amit predicted that simulation and reinforcement learning would become unquestionably central. Monica envisioned a marketplace of simulation environments that enable robots to continuously learn new contexts before deployment. Suraj pointed to deeper ecosystem collaboration and richer asset libraries, while Prasun emphasized scale—moving from millions to “billions of lines of simulation” to increase deployment confidence.

Stay Tuned for More LIVE Updates from CadenceLIVE Silicon Valley!