By Massimiliano “Max” Versace
AI executive, computational neuroscientist, and founder working across robotics, brain-inspired AI, and commercialization.
A robot can have a beautiful answer and still drop the glass.
That is the part of artificial intelligence that has kept my colleagues and me busy for much of our careers. Put intelligence inside a machine, give it sensors, ask it to do something useful, and the questions become wonderfully concrete. How quickly can it react? What happens when the light changes? Can it learn something new? How much battery is left?
Physics has a way of shortening the discussion.
Today, these questions sit at the center of Physical AI: artificial intelligence that helps machines perceive, learn, and act in the physical world. In robotics, physical intelligence brings together perception, learning, planning, control, sensing, and the hardware that makes action possible.
For the teams I have helped assemble and lead, those connections have been a practical concern for years—from computational neuroscience at Boston University and NASA-supported autonomous robotics research to commercial AI deployment and semiconductor work.
Looking back over roughly fifteen years of research, coding, talks, and reporting, one phrase keeps coming back: building brains for bots. It even became the title of an early interview about building an AI company.
Behind that phrase were neuroscientists, software engineers, roboticists, and commercial colleagues trying to make the pieces work together. Bringing those people together—and turning their different perspectives into a dependable product—became as important as the algorithms.
Physical AI in October 2026: new tools, familiar questions
The latest announcements make this a good moment to look back. On October 21, NVIDIA published “Cosmos Opens New Tools to Train, Evaluate and Deploy World Models”, describing tools for preparing data, adapting models, and evaluating physical AI. Its October 29 post, “How Developers Turn Ideas Into Simulations With Frontier AI Agents”, describes using AI agents with physics, rendering, and sensor-simulation tools.
These are meaningful developments. They also bring back a question our teams were asking years ago: can a machine acquire useful experience in a virtual world, then carry that learning into a physical one?
The October 17 article “From Computational Neuroscience to Physical AI” explored the connection between brain-inspired research and deployed intelligence. Here, the focus is the robotics lesson: why the technical continuity is exciting, and why a better artificial brain still needs a well-defined job.
The capabilities available today are substantially different. Many of the questions that determine whether they become useful products are familiar.
Computational neuroscience and the foundations of physical intelligence
The starting point was computational neuroscience. How do biological systems learn, retain useful information, and adapt when their environment changes? Those are good questions for neuroscientists. They become very practical questions when a robot encounters something its designers did not anticipate.
At Boston University, where I co-founded and directed the Neuromorphics Lab, our team explored ways to bring those ideas into artificial systems. BU’s account of the laboratory documents that work. Earlier research with Stephen Grossberg on learning in thalamocortical circuits provides part of its scientific foundation.
In 2010, Ben Chandler and I wrote “MoNETA: A Mind Made from Memristors” for IEEE Spectrum, describing brain-inspired computing in the context of the DARPA-supported work. The article brought together questions our colleagues were tackling across neuroscience and computing: how should the learning algorithm and the hardware executing it be designed together?
One lesson from that period was to start with the behavior the machine needs. Memory, sensing, learning, and computation all contribute to that behavior. Optimizing them separately can leave the system struggling when they finally have to work together.
The same applies to the people building it. A model that makes sense to a neuroscientist still needs to become software an engineer can run, on hardware a robot can carry.
Sim-to-real robotics: from virtual brains to physical machines
Our involvement in this line of research goes back to the DARPA SyNAPSE work that began in early 2009, a date documented in NASA’s account of the research. The public record of the team’s simulation-to-robot approach followed in the next few years.
The 2010 MoNETA article outlined a research roadmap: evaluate many simulated agents, identify useful architectures and behaviors, and bring the resulting intelligence into robotic platforms.
A particularly concrete account appeared in June 2011: “Transforming robotics with biologically inspired learning models,” by Aisha Sohail, provided by the National Science Foundation. It describes the lab developing artificial brains in virtual environments before deploying them in robots, testing many candidate brains in parallel, and using selected versions on an iRobot Create platform. Our later paper, “The Animat: New Frontiers in Whole Brain Modeling”, records another part of that collaborative research program.
Sim-to-real learning asks whether behavior developed in simulation transfers successfully to a physical machine. The tools have changed enormously. The engineering question is familiar: which parts of what a machine learns remain useful when it encounters real sensors, imperfect motors, and an environment we do not fully control?
Today’s world models and robotics learning systems offer capabilities and scale that were unavailable to us then. But the transition between simulated experience and physical behavior still has to be earned through testing.
Simulation lets a team explore more possibilities than a room full of physical robots. It also lets that team become very good at the wrong world if the simulator leaves out something important. A convincing image does not, by itself, establish correct friction, contact forces, sensor noise, or motor response.
Physical testing remains the examiner.
NASA, autonomous robotics, and intelligence at the edge
A machine exploring unfamiliar terrain cannot assume that every useful decision will arrive from somewhere else. It needs enough local intelligence to interpret what it senses and respond within the time available.
Working with NASA colleagues, our team explored autonomous navigation for planetary environments and subsequent applications on Earth. NASA describes this work in “A ‘Brain’ for Cars, Tested for Mars” and “From Mars Rovers to Factory Assembly Lines.”
The enduring lesson was that autonomy has to be designed around the conditions in which the machine will operate. Connectivity, available power, sensing, and response time shape what intelligence can actually do.
This is where Edge AI and Physical AI intersect. Edge AI concerns computation close to where data is generated. Physical AI concerns intelligence interacting with the physical world. A robot may use both local and remote computing, but its architecture must account for the decisions that cannot wait for a network connection.
Some of those shared lessons reached a wider audience through the 2014 TEDxFulbright talk, “Neuromorphics — Machines to Life,” and the 2016 NVIDIA GTC Europe presentation, “Brain-in-a-Box.” Putting intelligence onboard forced us to think about the complete machine: what it could sense, what it could compute, and how quickly it needed to respond.
Commercializing robotics AI: machines need a brain, and companies need customers
Building an AI company introduced a second education. A promising algorithm still needs an engineering team, a product, customers, partners, financing, and a reason for somebody to buy it again.
The company-building side of that work appeared in VentureBeat’s 2017 financing coverage and The Boston Globe’s reporting. Behind those milestones was the less visible work of deciding which applications to pursue and what customers actually needed.
The Unite.AI interview and conversation with the Center for Data Innovation discuss that progression from neural-network research into commercial AI, including continual learning and visual inspection.
For me, leading the company meant helping people with different expertise converge on the same customer problem. Research, engineering, product, and commercial teams each saw a different part of it. Progress depended on making those perspectives inform one another.
Choosing a first market determines which problems the team must solve now, which can wait, and what counts as success. An elegant demonstration and a dependable customer installation ask different things of a company.
The same is true in robotics today. A machine needs a useful job, an operating environment its builders understand, and an economic reason to exist. The people deploying it need to know how to teach it, maintain it, and recognize when it needs help.
Those questions belong in the architecture discussion from the beginning. They also belong in the boardroom.
Drones and telepresence: lessons from earlier robotics markets
Those of us who have worked through several waves of robotics enthusiasm recognize the moment when a promising capability becomes a prediction that machines will soon be everywhere. Our teams contributed some enthusiasm themselves. Read the early ambitions and you will find plenty of it.
Drones provide a useful example of uneven commercialization. NASA’s account connects our research to drone collision avoidance, and a later industrial drone-inspection collaboration described applying AI to inspection. Those are concrete tasks. The larger vision of autonomous flying machines becoming ordinary everyday products has followed a much less uniform path.
In August 2023, Skydio announced that it would sunset its consumer offerings and focus on enterprise and public-sector customers. Its explanation emphasized applications including inspection and public safety. The technology had useful capabilities; the company chose to concentrate on customers with different needs and economics.
Telepresence robotics offers another caution. The attraction was easy to understand: give a remote person a mobile physical presence. Yet a compelling demonstration does not settle how often mobility is needed, whether a customer will pay for it, or who maintains the machine. Bloomberg reported that Suitable Technologies, the company behind Beam, filed for bankruptcy in February 2020, after selling or leasing thousands of devices. Other telepresence products and applications continued, but the path to a sustainable business was clearly harder than the appeal of the demonstration suggested.
These examples reinforce a lesson that technical and commercial teams have to work through together: usefulness and a sustainable business must be established separately. For a drone inspection service, does the complete workflow produce a trustworthy, actionable inspection result? For a telepresence robot, does moving through the site deliver enough value beyond an ordinary video call to justify the cost and operational effort?
These are questions for the next robotics wave as well. More capable AI makes them more interesting. It does not make them disappear.
Deploying Physical AI starts with a specific problem
A general platform can be valuable. But “our robot can do many things” leaves the customer to work out which thing is worth paying for.
A better starting point is a precise promise: inspect this class of defect on this production line; move these parts between these stations; let this remote specialist examine this piece of equipment. Define the environment, required reliability, permitted human intervention, and cost. Then solve the workflow through to the point where the customer can depend on it.
“Fully solving” a problem does not mean promising perfection in every conceivable circumstance. It means delivering the agreed outcome within an explicit operating envelope, including exception handling, integration, maintenance, and support. A human fallback can be part of that solution if its frequency and cost make sense.
That focus unlocks deployment. Deployment can then generate experience, revenue, and evidence for expanding the product’s scope. Generality becomes something a team earns through demonstrated capability rather than an excuse to postpone finishing the first application.
Continual learning: a deployed robot will encounter something new
One of the recurring questions in our work was how to teach a neural network something new while preserving useful knowledge it already had. Our work on continual learning addressed incremental learning and the problem commonly called catastrophic forgetting.
For robotics, the practical motivation is easy to see. A useful machine has to work beyond the exact conditions of its demonstration. New objects arrive. Materials change. A customer’s environment is different from the laboratory.
Lessons from this work informed the RoboBusiness talk, “How to Train Your Robot’s AI,” and the IEEE Spectrum article “The Next Frontier in AI: Nothing.” Both explore questions about adaptation and recognizing the unfamiliar.
Physical intelligence will increasingly be judged by how well a system handles these moments. That includes knowing when to ask for help. Continual learning is a capability to engineer and validate carefully, especially when a machine’s actions have physical consequences.
From robotics research to industrial AI deployment
Industrial visual inspection may look less glamorous than a humanoid carrying a tray. It forces wonderfully serious questions about reliability, data, integration, and cost.
The connection to robotics is perception: a machine needs useful information about what is in front of it before it can make a useful decision. Inspection systems and mobile robots differ, but both expose the distance between a model performing well in a test and an AI product performing well in the world.
That broader transition was the subject of the Forbes Technology Council article “From Digital to Physical: The Ultimate Challenge for AI” in 2021, and later “Mastering AI Integration in Manufacturing: Unmasking the Myths.”
The work eventually moved beyond research prototypes into commercial applications. In its 2026 Space Technology Hall of Fame announcement, the Space Foundation reported approximately 80 million deployments of the technology across applications.
That was a collective achievement, built through the contributions of colleagues, partners, and customers over many years. Progress depended on finding applications where the technology solved a sufficiently important problem, then doing the engineering required to make it dependable.
Physical intelligence connects sensors, semiconductors, and robot behavior
Working today with colleagues in the semiconductor industry brings sensing, signal processing, memory, and efficient computation into the same design discussion.
The next generation of robotics requires that view across the entire system: how a signal is acquired, how information is represented, where memory and computation sit, and how the result becomes an action.
Consider a robot hand picking up a glass. The task involves vision, touch, force, motion, and timely correction. A useful intelligence architecture has to account for that interaction. Model capability matters enormously, and so do the measurements and timing that allow the machine to use it.
These questions also framed the discussion on “The Physical AI Stack” at Imagination in Action’s Minds in Motion event, where I joined fellow researchers and industry leaders at MIT’s Samberg Conference Center on October 30, 2026.
For organizations building in this area, the opportunity is to bring disciplines together early. Robotics engineers, AI researchers, semiconductor designers, and product leaders need to agree on what the complete machine must accomplish.
That agreement takes work. It requires people who can challenge one another’s assumptions, explain constraints outside their own discipline, and stay focused on the behavior the customer needs.
What fifteen years of building brains for bots has taught us
Start with a useful behavior. Understand the environment. Design the sensing and computation around the task. Give learning a clear purpose. Measure the complete system. Build a business that can support it after the demonstration ends.
These sound like simple instructions. Anyone who has tried to follow them knows how much work fits inside each sentence.
Working alongside people who could take an idea from a neural model to a deployed product has made me optimistic about physical intelligence, and demanding about what we should expect from it. The challenges are substantial, but many are recognizable. They become more tractable when the right people work on a clearly defined problem and test their assumptions against the world.
The original ambition still holds: build a brain for a bot, and give it something useful to do.
Preferably without dropping the glass.
About Max Versace: Physical AI, robotics, and commercialization
Massimiliano “Max” Versace is an AI executive, computational neuroscientist, entrepreneur, and former CEO whose work connects Physical AI, robotics, neuromorphic computing, Edge AI, and semiconductor technology. He co-founded and directed Boston University’s Neuromorphics Lab and has assembled and led multidisciplinary teams working across brain-inspired research, autonomous systems, and commercial AI deployment.
His experience includes NASA- and DARPA-supported research, nearly two decades of founder/CEO leadership, and current leadership of Emergent AI work in the semiconductor industry. He is a Board Advisor to the EDGE AI FOUNDATION.
Further background: Max Versace’s executive profile, computational neuroscience and Physical AI career overview, and talks, interviews, and media coverage.
Max Versace’s talks, research, and reporting on AI and robotics
The following archive brings together the research, conversations, and reporting behind this article. Links use descriptive labels; historical titles and affiliations appear on the original pages. Some publishers require a subscription, and some recordings are presented through an accompanying page on this website.
Robotics, NASA, and independent reporting
- IEEE Spectrum — Deep Learning AI for NASA Powers Earth Robots
- National Science Foundation account — Transforming robotics with biologically inspired learning models
- NASA Spinoff — Planet-Navigating AI “Brain” Helps Drones and Cars Avoid Collisions
- NASA Spinoff — From Mars Rovers to Factory Assembly Lines
- NASA — A “Brain” for Cars, Tested for Mars
- NASA — From Mars Rovers to Factory Assembly Lines
- VentureBeat — Financing the development of AI for autonomous machines, 2017
- The Boston Globe — Financing AI software development, 2017
- The Boston Globe — Mobile AI applications, 2017
- Fortune — AI entrepreneurship and financing, 2017
- TechCrunch — Bringing machine learning to the edge, 2017
- Il Sole 24 Ore — Il ratto nel robot, 2011
- TIME — Machine-Learning Programs Help Doctors and Their Patients
- Associated Press — Decade Since Recession: Thriving Cities Leave Others Behind
- Inc. — This Artificial Intelligence Expert Breaks Down Your Burning Questions
Talks and video recordings
- TEDxFulbright — Neuromorphics: Machines to Life
- NVIDIA GTC Europe — Brain-in-a-Box
- MIT Technology Review Future Compute — How AI Is Changing the Way We Work
- MIT Future Compute — accompanying retrospective
- Fulbright Association — Humanizing AI: From Myth to Reality
- Fulbright Association — conference speaker listing
- AI, Immigration, and Entrepreneurship — video interview
- AI, Immigration, and Entrepreneurship — accompanying account
- RoboBusiness — How to Train Your Robot’s AI
- TEDxCortina — Unlocking the AI Mirror: What Makes Us Unique?
- Robotics Summit & Expo — speaker profile
- Minds in Motion — The Physical AI Stack, October 2026
Interviews and podcasts
- The VentureFizz Podcast — Building and leading an AI company
- Unite.AI — Neuroscience, continual learning, and entrepreneurship
- Tech Talks Daily — AI in Manufacturing: A Second Set of Eyes
- Association for Advancing Automation — Utilizing Vision AI in Manufacturing
- Center for Data Innovation — 5 Q’s With Max Versace
- Immigrant Nation — Replicating the Human Brain: The Future of AI
- Immigrant Nation Podcast — accompanying post
- The Early Stage — Building Brains for Bots
Scientific research and technical writing
- Spikes, synchrony, and attentive learning by laminar thalamocortical circuits
- Brain Research paper — DOI
- MoNETA: A Mind Made from Memristors
- An Introduction to Silicon Brains: Dr. Massimiliano Versace
- From Synapses to Circuitry: Using Memristive Memory to Explore the Electronic Brain
- The Animat: New Frontiers in Whole Brain Modeling
- Does Artificial Intelligence Require Specialized Processors?
- The Next Frontier in AI: Nothing
Forbes Technology Council articles and contributions
- Mastering AI Integration in Manufacturing: Unmasking the Myths
- Software to Hardware: How to Shift to Building Physical Tech — expert-panel contribution.
- Manufacturing, Sustainability and Profitability: How AI Can Make Us Greener
- “I Doubt, Therefore I Am,” Said AI
- Why 2022 Will Be an AI-Powered Renaissance
- How AI and Automation Can Address America’s Broken Supply Chain
- From One of a Million to One in a Million: The Path to Customization Passes Through AI and Automation
- How Manufacturers Have Turned the Table on Automation
- From Digital to Physical: The Ultimate Challenge for AI
- Why “Explainability” Is the Next Step for AI in Manufacturing
- Rebounding From the Pandemic: What Our Past Teaches Us About the Future
- The Coming of Age of Industry 4.0 in a Post-Pandemic World
- From Computer Vision to Deep Learning: How AI Is Augmenting Manufacturing
- 2020 and the Dawn of AI Learning at the Edge
- 2020 AI Predictions: What We Got Right in 2019 and What’s in Store for 2020
- The World Needs WordPress for AI
- The Next-Generation AI Brain: How AI Is Becoming More Human
- If You’re Worried About AI Taking Your Job, Consider This
Institutional profiles and current work
- Physical Intelligence and Emergent AI — semiconductor industry perspective
- EDGE AI FOUNDATION — Board of Directors and Advisors
- Boston University — Pushing the Boundaries of Artificial Intelligence
- Boston University — The Robot as Decider
- Space Foundation — March 2026 Space Technology Hall of Fame announcement
- EDGE AI FOUNDATION — Partnership supporting sensing and intelligent edge computing
- Industrial Ethernet Book — Looking Ahead to 2026: AI Predictions
Further background
- About Massimiliano “Max” Versace
- Executive profile: AI leadership and commercialization
- Media & Speaking: talks, interviews, and press