Digital Dialogs (Episode 4 | S5) - Spatial Intelligence & Scene Graphs

In this episode, Ilario Gerlero joins Digital Dialogs to explore the rapidly expanding field of edge AI and on-device intelligence. The conversation covers what it means to run AI directly on hardware at the point of action, the engineering challenges of deploying capable models on constrained devices, and where this paradigm shift is heading across industries. Listeners will gain insight into why moving intelligence to the edge is becoming essential for real-time, private, and resilient AI systems — and what it takes to make it work in practice.

In this episode, Brigitta Errera and Vincenzo Miccichè from Reply's Area42 lab join Digital Dialogs to share how they're tackling sim-to-real transfer for humanoid robotics — using Gaussian splatting to build high-fidelity simulated environments, training policies in sim, and deploying them on physical robots. The conversation also explores how agentic coding frameworks like Code as Policies are creating new pathways for language-driven robot programming, and how these two approaches might converge. Listeners will get a hands-on view of what it takes to bridge the reality gap from inside a lab that's actively doing it.

In this episode, Anna Rossebastiano and Alessio Magurno join Digital Dialogs to explore how teleoperation bridges human perception and remote robot action. The conversation covers how operators perceive and control distant environments in real time, the system architecture that makes responsive teleoperation possible, and the emerging role of shared autonomy in reducing operator burden. The episode also dives into how teleoperation itself becomes a data collection tool — capturing egocentric video, joint positions, and motor commands that can be used to train autonomous policies. Listeners will learn how the path from human-guided control to fully autonomous robot behavior is shorter than it might seem.

In this episode, Maxim Stoljar joins Digital Dialogs to explore how digital twins and NVIDIA Omniverse are transforming the way organizations simulate, test, and optimize complex real-world systems. Maxim and his team have been working hands-on with Omniverse — including presenting their work at NVIDIA GTC — and have delivered real client projects such as a warehouse automation digital twin for Otto Group. The conversation covers what sets the current generation of digital twins apart, the practical challenges of building physically accurate simulations in production, and where this technology is heading next. Listeners will gain insight into why digital twins are becoming a foundational layer for Physical AI, and what it takes to move from concept to production-grade simulation at scale.

In this episode, Perumal Kumaresa explores why synthetic data is becoming one of the most important enablers of Physical AI. As organizations try to build systems that can perceive, predict, and act in the real world, they run into a basic problem: the most valuable training data is often the hardest, riskiest, or most expensive to collect. Perumal explains where synthetic data can clearly outperform real-world collection, especially for rare failures, safety-critical scenarios, privacy-sensitive environments, and edge cases that teams may never capture often enough in production. The conversation then turns to the harder question of trust: how do you validate that a model trained in synthetic environments will actually hold up when deployed in the messiness of the real world? We also widen the lens beyond physical systems to discuss synthetic audiences and synthetic A/B testing, and whether these are a natural extension of the same idea or a different category with different risks. Finally, Perumal looks ahead to a possible synthetic-first future, where teams begin in simulation, generate targeted scenarios at scale, and only then fine-tune in reality—potentially changing the speed, cost, and ambition of what companies can build.

In this episode, Michael Cruz joins Digital Dialogs to explore world models — AI systems that learn internal representations of how the physical world works. The conversation covers what world models are, why current generative AI falls short in understanding physical reality, the technical challenges of teaching machines to simulate physics, causality, and spatial dynamics, and where these capabilities are heading across robotics, autonomous systems, and simulation. Listeners will gain insight into why building AI that truly understands the physical world is one of the field's most important open problems, and what it will take to get there at scale.

In this episode, Ana-Maria Lacatusu explores the concept of "Spatial Intelligence" and why it serves as the crucial missing link for autonomous operations. She breaks down complex concepts like scene graphs, explaining why it's not enough for autonomous robots or smart buildings to simply recognize objects—they must understand the relationships between them. Ana-Maria discusses the tangible business value of having a digital, structured understanding of the environment for companies with physical operations, such as warehouses, construction sites, and hospitals, highlighting how turning physical space into structured data drives operational efficiency and automation. Finally, the conversation looks toward the future of embodied AI, exploring emerging capabilities that will allow systems to navigate and collaborate safely alongside humans in complex, unstructured environments without constant supervision. This episode reveals how turning physical space into structured data is paving the way for the next generation of intelligent, autonomous systems.

In this episode, Dr. Dimitris Vlitas unveils "Decoding Antiquity," one of the most compelling examples of AI meeting the physical world. In collaboration with the Austrian Academy of Sciences and Mistral AI, his team has produced the world's first multimodal large language model for Ancient Greek, trained on the largest digital corpus of historical Greek ever compiled—over 600 million words spanning ancient, medieval, and early modern periods. With fewer than 12% of the estimated 500,000+ surviving Ancient Greek papyri ever transcribed, and the rest requiring rare specialist expertise to reconstruct from fragmented physical artifacts, this AI system can perceive damaged documents, interpret fragmentary text, and generate the most likely complete versions for scholarly review. Dimitris walks through the extraordinary technical challenges—from curating chronologically distinct training datasets spanning millennia, to designing evaluation processes that combine traditional AI metrics with direct scholar feedback, to the multimodal pipeline that turns a photograph of a two-thousand-year-old papyrus into readable text. The conversation broadens to explore the strategic architecture behind the project: a custom foundation model deployed on Mistral's EU-sovereign platform where background IP remains the client's property and outputs create new, defensible competitive advantage. This episode sets the stage for Season 5 by demonstrating that Physical AI starts with teaching machines to truly see and understand the physical world around us—whether that's an ancient manuscript or a modern factory floor.

In this episode, Johannes Morzeck explores how Vision Foundation Models (VLMs) and specialized industrial models are fundamentally transforming computer vision from expensive custom R&D projects into scalable, ready-to-use business solutions. Johannes discusses the real breakthroughs compared to classic computer vision and the practical challenges organizations face when moving these massive models from proof-of-concept to production in real-world environments like factories and retail stores. The conversation delves into the complexities of balancing model performance with the cost of running them at the edge. Looking toward the future, Johannes shares his insights on how these models will reshape industries reliant on visual data, envisioning a world where cameras serve as "universal sensors" capable of answering any question about physical operations without the need for custom training.

In this season opener, Rene introduces the concept of Physical AI—the next frontier where digital intelligence meets the real world. Moving beyond generative AI on screens, this episode explores how Spatial Intelligence allows machines to truly understand their environment, not just capture pixels. Rene breaks down the core architecture of these systems, explaining the role of World Models in helping devices learn the laws of physics and predict outcomes in dynamic environments. The discussion covers the critical pipeline of Physical AI: from Perception using advanced computer vision and sensors, to Simulation via digital twins and synthetic data for safe training, and finally to Action through embodied agents and robotics. Rene sets the stage for Season 5 by highlighting key use cases in industry, logistics, and healthcare, and outlining how this technology is shifting automation from pre-programmed repetition to adaptive, autonomous intelligence.