Physical AI is reshaping robotics from fragile scripts into resilient collaborators capable of real-world judgment and dynamic decision-making. A new generation of automation platforms is closing the gap between perception and action, enabling machines to understand, adapt and improve in complex industrial environments.
The story of automation has long been one of rigid control and fixed boundaries. Machines were expected to repeat, not think. Scripts dictated every movement, and deviation meant failure. That paradigm is shifting. Physical AI, the convergence of artificial intelligence and robotics in real-world environments, is bringing a new layer of perception, adaptability, and cognition to machines that, until now, were simply dumb muscles. At NVIDIA GTC 2025, a panel of leading technologists and robotics experts laid out a compelling vision of what comes next as industrial automation enters its intelligent phase.
The shift is not abstract. Across manufacturing, logistics, and warehousing, AI-driven systems are already augmenting how machines see, move, and make decisions. These are not concepts in development; they are deployed solutions tackling precision assembly, visual inspection, and multi-robot coordination. At the heart of this transformation is a growing ability to make sense of physical environments, recognise, interpret, and adapt to them in real time.
From robot operators to robotic operators
According to Madison Huang, Director of Product and Technical Marketing, Physical AI Platforms, NVIDIA, the turning point is a full-stack approach that unifies software, hardware, infrastructure and development frameworks. NVIDIA’s AI Accelerator platform already enables partners to build next-generation robotic solutions, whether T-Robotics interpreting natural language commands or IKA using reinforcement learning to solve complex assembly tasks involving physical contact. “Our goal is to accelerate time to market for developers and allow applications to evolve with business needs,” Huang explains. “This is not just about enabling innovation; it is about making it deployable, scalable, and valuable across the industrial landscape.”
For Huang, the real advantage lies in combining accessibility with deep technical capability. The success of Universal Robots, which has sold more than 100,000 collaborative robots, proves the value of user-centric design when paired with powerful AI ecosystems. “The key to all this is platforms and partnerships,” she adds. “Physical AI is no longer a concept for tomorrow; it is a reality today.”
At industrial automation company, Vention, Etienne Lacroix describes a similar journey. Rather than treating AI as an add-on, his team has embedded it into an automation platform that mirrors the plug-and-play nature of Lego. By unifying mechanical design, PLC logic, and robot programming within a single ecosystem, the company has drastically lowered the barrier to automation. “We are now seeing a new design paradigm,” Lacroix explains. “Simpler arms, fewer jigs, fewer sensors, and more intelligence packed into lower-cost components.”
The company’s AI operator technology already performs unstructured 3D picking tasks with over 90 per cent reliability and inference times under five seconds. Powered by MachineMotion AI, these systems are not hypothetical prototypes. They are delivering an average payback period of 1.3 years, and in the US, closer to 1.1. The implication is clear: intelligent robotics is becoming cost-effective at scale, not just in theory but in practice.
Digital twins and simulation as scaffolding
The future of intelligent robotics is not simply about machines learning on the fly. It is about learning in silico before they ever reach the shop floor. At Siemens, Dr Eugen Solowjow has spent years examining the practical hurdles of deploying AI systems that deal with infinite real-world variability. The solution, he believes, is simulation.
Using a combination of Siemens Process Simulate, NVIDIA Omniverse, and the company’s PLC stack, Solowjow’s team tests robotic behaviours in richly modelled digital environments before deploying them in physical settings. This approach has already delivered results with Semantic Robot Pick AI, a foundation model for pick-and-place tasks agnostic to robot and camera types. “Simulation allows us to test foundation models in context before they are deployed in real operations,” he explains. “That reusability across applications is where we see massive value.”
This simulation-to-reality pipeline enables companies to validate AI systems against edge cases and production variability without shutting down their lines. These workflows must become standard for robotics to achieve widespread adoption, particularly among small and medium manufacturers. As Solowjow puts it, customers are not interested in models or datasets; they want plug-and-play reliability with clear ROI.
Intelligent vision and rapid learning
One of the biggest barriers to automation has always been perception. Robots struggle to operate in unpredictable conditions, where lighting, reflection, and object irregularity often derail traditional vision systems. At Solomon, Johnny Chen believes that AI-powered 3D vision is finally closing that gap.
“Most robots still rely heavily on pre-programming and cannot adapt easily to new tasks or changing environments,” he explains. “We developed our own systems that see, understand, and adapt.” Solomon’s platforms support a range of camera types and use few-shot learning to train models in under 30 seconds from just two images. This flexibility makes them suitable for various applications, from semiconductor defect inspection to real-time medical device analysis.
Using NVIDIA Isaac Motion, Solomon has accelerated motion planning by 80 times, while foundation models like Amrita’s Pulse eliminate the need for time-consuming calibration by matching CAD data with point clouds in real time. “Future learning, vision-language models, and superhuman vision allow humanoids to identify, inspect, and interact with new objects faster and more accurately than humans,” Chen explains. “They can monitor machines around the clock, scan barcodes instantly, and even locate contact lenses from eight feet away.”
Physical context as a competitive edge
Physical AI is distinct from other branches of AI because it operates within real-world constraints and uses them as inputs. Amir Bousani of RGo Robotics calls this the “information layer,” a data set that includes object locations and metadata, spatial logic, workflow priorities, and machine behaviours.
This layer underpins simulation and real-time control, allowing generative agents to direct entire fleets of machines or respond to queries like, ‘Show me unsafely stacked objects’. The system can explain its reasoning visually and recommend actions, a form of transparency rarely seen in automation. “The full loop, from real-world data to simulation, decision-making, and action. Is where physical AI shines,” Bousani explains.
He describes digital twins not as passive visualisations but as active, continuously updated models that reflect site changes and allow for predictive adjustments. When AI has full physical context, it no longer reacts and reasons.
Beyond discrete deployments
At Teradyne, James Davidson sees a future where intelligent automation is not a one-off investment but a continuous capability. By unifying perception, planning and control, his team has created systems that move from explicit step-by-step programming to more fluid, abstracted logic. Operators can now guide robots using demonstrations or natural language, reducing the need for specialist intervention.
This modularity and interoperability are essential for scaling. Davidson points out that most factories use a patchwork of legacy systems and vendor-specific protocols. AI solutions must be hardware-agnostic and easily integrable. “The challenge is to ensure that the system can adapt not just to one environment but to many,” he explains. “It needs to be robust to variations in lighting, clutter, object orientation, and human interaction.”
Simulation remains key, not just for prototyping but also for continual learning. Foundation models trained in digital twins can be deployed with confidence and adapted post-launch through real-world feedback. The goal is not just intelligence at the edge but resilience throughout the lifecycle.
The industrial world is not entering a post-human phase of automation. Rather, it is entering a post-script phase, where robots do not simply execute code but participate in decision-making. That shift is not philosophical. It is infrastructural, economic, and already underway.
