Max Versace

AI, academia => real-world

Physical AI: From Brains to Machines

Illustration of a biological neuron connecting through a neural network to a robotic hand gripping a metal component.

By Massimiliano “Max” Versace
Computational neuroscientist, AI executive, inventor, and entrepreneur.

Ask a robot to pick up an espresso cup and a surprising amount of science turns up. Vision has to locate it. Touch has to detect contact. Motors have to apply enough force to hold it, preferably without turning it into a ceramics experiment. Somewhere in that process, a computer has to decide what to do next, before the coffee ends up on the floor.

Physical AI is artificial intelligence that uses sensory information to perceive, predict, learn, and guide action in the physical world. It connects neural computation with sensing, memory, planning, control, and the hardware that makes behavior possible. Robotics is its most visible expression, but the same questions arise in industrial inspection, autonomous vehicles, medical devices, and intelligent sensors.

This page brings together the ideas, research, patents, articles, and outside reporting behind my work in this field. My colleagues and I have approached these questions through computational neuroscience, DARPA-supported neuromorphic research, NASA-supported robotics, and the decidedly educational experience of turning research into products. The terminology has changed over the years. The requirement that the machine actually work has been remarkably consistent.

Recognition and outside reporting · Definitions · The Physical AI stack · Neuroscience and neuromorphic computing · Sim2Real · Edge AI and continual learning · Deployment · Start reading · Publication archive · Patents · Questions

Recognition and reporting beyond this website

The most useful account of a career is one other people can check. These sources document the work, the teams behind it, and how it moved beyond the laboratory.

  • Space Technology Hall of Fame, 2026. The Space Foundation inducted the AI brain our team built, recognizing our company and NASA Langley Research Center, with individual commendations for Heather Ames, Massimiliano Versace, Anatoli Gorchet, and Mark Motter. Its announcement describes simulation-based terrain learning, transfer to physical robots, and subsequent commercialization. It reports deployment in approximately 80 million systems. That figure covers the technology across applications; it does not mean 80 million autonomous robots.
  • NASA technology transfer. NASA’s 2018 Spinoff account traces the work from DARPA SyNAPSE to NASA-supported autonomous systems. NASA’s 2024 account follows the technology into industrial visual inspection. These were research and technology-transfer programs, not a claim that our software flew on a Mars rover.
  • Netexplo recognition, 2018. UNESCO’s account of the Netexplo Innovation Forum lists our collaboration with Air Shepherd among the award-winning applications, for work on detecting poaching. This recognition concerns the application and the collaborating organizations.
  • Independent technical and business reporting. IEEE Spectrum examined the transition from NASA research to onboard robot intelligence. TechCrunch, VentureBeat, and The Boston Globe covered the company’s development and financing. The archive below includes the original articles.
  • Research and professional leadership. Boston University documented the Neuromorphics Lab and my role as its director. Today, the EDGE AI FOUNDATION lists me as a Board Advisor; its partnership announcement identifies my role as Vice President of Emergent AI at a global semiconductor company.

These are shared achievements. They required researchers, engineers, partners, customers, and colleagues willing to stay with a problem after its first impressive demonstration.

Matt Luciw, NASA Langley engineer Mark Motter, and Max Versace standing together outside a building.
The people behind the NASA collaboration. Matt Luciw, Mark Motter of NASA Langley, and Max Versace, from left to right. NASA’s account traces the STTR-supported work from brain-inspired navigation research toward commercial applications. Image source: NASA Spinoff (2018).

Further recognition of the teams and technology

These recognitions belong to the teams, companies, and technologies named in their original records. They are part of the history of my work, with credit shared accordingly.

Physical AI, embodied intelligence, and robotics: what do the terms mean?

I use Physical AI to describe intelligence that has to account for the consequences of operating in the physical world. A useful definition needs to tell an engineer what must be built and measured. Merely putting a language model inside a robot does not settle the matter.

TermWhat it describesEngineering question
Physical AIAI used to interpret physical conditions and guide decisions or actions in physical systems.Does the complete system achieve its task under real conditions?
Physical intelligenceThe ability to use sensing, prediction, and action to handle the demands of a physical environment.Does the intelligence account for the body, its sensors, and the consequences of acting?
Embodied intelligenceIntelligence shaped by interaction among an agent, its body, and its environment.What can the agent learn through its own actions and their sensory consequences?
RoboticsThe design and operation of machines that sense and act.How do mechanics, sensing, computation, and control produce dependable behavior?
Computational neuroscienceMathematical and computational models used to explain nervous-system function and behavior.Which mechanisms explain learning, perception, attention, and action?
Neuromorphic computingComputing architectures that draw on nervous-system organization or dynamics, often using distributed state and event-driven processing.Which biological principles improve the implementation of this workload?
Sim2RealTransferring models, representations, or behavior developed in simulation to physical hardware.Which assumptions survive the move to real sensors, materials, and actuators?
Edge AIAI computation close to the source of data or the point of action.Can it meet the local limits on latency, power, memory, connectivity, and cost?

Physical intelligence: what the machine has to understand

Physical intelligence is a useful way to describe the capability we are asking for. Physical AI describes the artificial systems we build to pursue it. The two expressions often overlap in robotics, but a useful discussion has to get past the terminology. The machine needs information about its surroundings, a way to anticipate consequences, and feedback that tells it whether an action worked.

Picking up a cup makes the distinction concrete. Recognizing the cup is one task. Estimating where to grasp it, responding when it slips, and adjusting force without breaking it require sensing, prediction, and control to work together. This is where computational neuroscience, embodied intelligence, and engineering meet.

These areas overlap, but they answer different questions. An inspection camera can use Edge AI without controlling a robot. A robot can combine local control with remote planning. A brain-inspired algorithm can run on a conventional processor. Choosing an architecture requires more precision than choosing a label.

The Physical AI stack: from a measurement to a useful action

The stack I care about begins before the neural network. A sensor measures something through noise, limited bandwidth, calibration error, and a particular point of view. Whatever the model does next depends on the information that survives that process.

  1. Sensing and signal processing. Cameras, microphones, tactile sensors, force sensors, encoders, and inertial sensors provide different observations. Sampling, synchronization, filtering, and calibration determine whether those observations can usefully be combined.
  2. Perception and state estimation. The system estimates what is present, where it is, how it is moving, and how uncertain those estimates are. A classification score alone is not a complete description of the world.
  3. Memory and prediction. An internal model retains useful history and predicts what may happen. For a robot inserting a connector, the relevant prediction may be contact and resistance, rather than every pixel in the next camera frame.
  4. Planning and action selection. The system chooses an action that advances the task within its operating limits. Modern vision-language-action models can connect visual observations and instructions to robot actions; RT-2 is one published example. Its reported generalization results should not be mistaken for a guarantee across arbitrary tasks and machines.
  5. Control and actuation. Commands become forces and motion. Feedback must correct errors on the timescale of the task. A slower planning process can coexist with a much faster local control loop.
  6. Learning and supervision. Experience can improve the system, provided changes are evaluated, recorded, and constrained. Failure detection, human intervention, and recovery belong in the design.

Consider timing. End-to-end delay includes sensing, transport, computation, scheduling, and actuation. A fast neural network can still sit inside a slow machine. The same applies to energy: measuring the accelerator alone leaves out sensors, memory traffic, communication, and the motors doing the work.

The practical test is simple to state and difficult to pass: can the machine complete the task repeatedly, within the agreed time and energy budgets, including the awkward cases?

Physical AI is a feedback loopSensors observe the world. State estimation and prediction inform planning. Control and actuators change the world, producing new observations. Learning updates models under validation and supervision.Physical AI is a feedback loopPerformance belongs to the complete machine, not just the model.SensorsObserve and synchronizeState and predictionEstimate, remember, anticipatePlanningChoose a feasible actionControl and actuationApply forces and correct errorsPhysical worldContact, motion, consequencesLearning and supervisionValidate changes · detect failure · recoverMeasure latency, energy, uncertainty, task success, and human intervention.
Conceptual system diagram. Learning may update several components; deployment still requires validation on hardware.
CoCoRo cognitive cycle linking executive goals, sensory input, cognitive processing and motor commands through working memory.
From a goal to an action, and back again. Our CoCoRo controller organized sensing, cognition, and movement around a shared working memory. Figure 1, Galbraith, Guenther and Versace (2015), Frontiers in Neurorobotics 9:7. © 2015 the authors. CC BY 4.0. Figure content unchanged; converted to JPEG.

What computational neuroscience contributes

Biology remains an excellent source of uncomfortable questions for computer engineers. Animals combine perception, memory, motivation, and movement while continually encountering things they have never seen before. They do this with hardware that was not designed around a benchmark leaderboard.

My work with Stephen Grossberg on spikes, synchrony, and attentive learning in thalamocortical circuits examined how neural dynamics support attention and learning. Work with Jasmin Léveillé and Stephen on spiking dynamics and perceptual grouping considered how interactions across a network resolve ambiguous visual information. These are computational models of particular mechanisms, not complete explanations of intelligence.

For engineering, the attraction is in the questions they make us ask. How does feedback alter perception? Which information deserves attention? How can a system learn without erasing useful knowledge? What should happen when its expectations are wrong?

Neuromorphic computing: the physical implementation matters

At Boston University, our colleagues worked with Hewlett-Packard researchers on models and computing architectures in the DARPA SyNAPSE program. Ben Chandler and I described the effort in MoNETA: A Mind Made from Memristors. The collaborative paper From Synapses to Circuitry examined memristive memory and its relationship to computation.

Moving data has a cost. Keeping state closer to computation, communicating only useful changes, and exploiting sparse activity can matter as much as counting arithmetic operations. Whether a neuromorphic implementation wins depends on the workload, precision, hardware, and supporting software. A spike is not automatically an energy saving; the complete implementation has to earn that claim.

My perspective on that relationship is developed in The next AI revolution will be hardware + software. The aim is to learn from biology while remaining very practical about the machine being built.

Two stylized profiles linked by crossing pathways, illustrating the MoNETA feature in IEEE Spectrum.
MoNETA in IEEE Spectrum. Illustration by Chad Hagen for our article, MoNETA: A Mind Made from Memristors.
Cog Ex Machina architecture connecting a brain model, virtual or physical body, Cog framework, accelerator nodes and a photonic network.
Connecting a brain model to a body. Cog Ex Machina platform, Figure 1 in Snider et al., From Synapses to Circuitry, IEEE Computer (2011). © 2011 IEEE. Extracted diagram; content unchanged.

Sim2Real: a virtual world is a useful teacher, and an imperfect one

Our involvement in DARPA SyNAPSE began in early 2009, as NASA’s history records. By 2011, an NSF-provided account by Aisha Sohail described our lab creating artificial brains in virtual environments before deploying them in robots. The Animat, published in 2012, documents the broader collaboration. Those dates establish the history of our work; they are not a claim that our team invented simulation-to-real transfer.

Simulation gives a team repeatable experiments, faster exploration, and access to conditions that would be expensive or dangerous to create physically. The trouble starts when the robot learns a simulator’s convenient mistakes. Friction, backlash, compliance, sensor noise, lighting, contact, and communication delays all have ways of becoming important at precisely the wrong moment.

Domain randomization deliberately varies aspects of the simulated environment so a policy or perception model encounters a range of conditions. System identification, real-world calibration, and carefully designed adaptation can also reduce the gap. None removes the need to test on hardware.

World models add another tool: learned representations of how an environment evolves, potentially conditioned on actions. Their usefulness depends on the accuracy of the consequences that matter for the task. A plausible video of a successful grasp is not evidence that the corresponding forces are possible.

A sound evaluation holds out objects and conditions, measures performance on the physical machine, and records failures and human interventions. It also asks whether success survives changes in payload, wear, lighting, and sensor calibration. Reality is quite good at finding the variable we forgot.

The physical Calliope mobile robot with arm and depth camera, left, beside its Webots simulation, right.
The robot on both sides of the screen. Calliope in the laboratory (left) and its Webots counterpart (right), used in our study of exploratory learning and motor planning. Figure 3, Galbraith, Guenther and Versace (2015), Frontiers in Neurorobotics 9:7. © 2015 the authors. CC BY 4.0. Figure content unchanged; converted to JPEG.

Edge AI and continual learning

For a physical machine, compute location is a consequence of its constraints. Some decisions can tolerate a network round trip. Others cannot. This is the argument behind Edge AI is not a place, it’s a constraint.

There is another distinction that matters: edge inference runs a trained model locally; edge learning changes what the system knows using new experience. A device can do the first without doing the second. Continual learning adds the problem of acquiring new knowledge while retaining useful previous knowledge, within bounded resources.

Our work on Lifelong-DNN grew from these practical concerns. The Space Foundation’s account describes learning on edge devices without relying on cloud connectivity. The linked patent records and articles below provide more detail about particular methods. Different forms of incremental learning solve different problems; learning new visual categories is not the same achievement as learning an unrestricted new robot skill.

Adaptation also creates an obligation. A system that changes after deployment must be monitored for regressions, unexpected behavior, and forgetting. Data quality, validation, versioning, and rollback are part of learning in a product. An industrial customer should not discover that yesterday’s improvement broke today’s production run.

Boston street scene with colored detection boxes identifying pedestrians, cyclists and vehicles.
From research to perception on a city street. This example from our team’s software identifies road users and vehicles in a Boston street scene. NASA published it while describing the transition from its research support to commercial AI. The image illustrates object detection, one component of a driving system. Image source: NASA Spinoff (2018).

From a working model to a machine people can depend on

Leading an AI company taught me that the algorithm is only one part of what a customer buys. The rest includes installation, integration, usability, support, and an answer to the reasonable question: what happens when it fails?

ApplicationWhat must workWhat to measure
Industrial visual inspectionDetection under production variation, with useful operator feedback.Missed defects, false rejects, inspection time, changeover effort, and downtime.
Mobile robots and dronesPerception and navigation despite changing conditions and limited connectivity.Task completion, interventions, recovery, energy use, and performance inside the defined operating area.
Dexterous manipulationVision, touch, force, and motion working together through contact.Successful grasps, slip, damage, cycle time, and robustness across materials and objects.
Intelligent sensorsUseful local interpretation of noisy physical signals.Detection quality, end-to-end latency, energy, memory, and calibration stability.

These are evaluation questions, not claims that one architecture solves all four categories. Define the job, specify the conditions, measure the complete workflow, and count the human help honestly. A narrow application that works reliably can be an excellent business. A machine with an impressive list of possible applications can still have no customer.

The Center for Data Innovation interview discusses that transition from research to industrial AI. NASA’s account of factory inspection documents one concrete result. The scientific, technical, and commercial decisions have to inform one another from the beginning.

Reconstructed view from Calliope as its gripper lifts a green object, with position estimates and motor values.
Seeing the task from the robot. A reconstructed view during Calliope’s grasp-and-lift experiment. Figure 12, Galbraith, Guenther and Versace (2015). © the authors, CC BY 4.0. JPEG conversion.
A quadcopter beside an electrical transmission tower.
AI for infrastructure inspection. Image accompanying the 2020 drone inspection report in my article archive. The report describes applying vision AI to inspection workflows.

Start reading: Physical AI in practice

Publication archive: research, outside reporting, articles, and talks

This archive follows more than fifteen years of work, with a few earlier papers that help explain its foundations. Reporting by others, institutional accounts, my own articles, collaborative papers, and recordings are identified separately. Historical affiliations describe the role held at publication. Some publishers require a subscription. An older prediction is part of the historical record, not a statement of what has since been proved.

Independent reporting and NASA / NSF accounts

Talks and video recordings

Interviews and podcasts

Scientific research and technical writing

Forbes Technology Council articles and contributions

My contributor articles and expert-panel contributions; these are distinct from independent reporting about my work.

Institutional profiles and current work

Additional research records

More third-party coverage and recognition

Medium and related publications: my writing

LinkedIn publications and posts

More writing: my Medium archive, LinkedIn, and the Forbes Technology Council contributor index. These indexes help track additions and changes to older publisher URLs.

Further reporting and research

Patents: selected technical records

Patent records describe inventions and their claims. They are not independent benchmarks of a deployed system. The records below name me among the inventors; several reflect joint work. Grant and application numbers are distinguished, and current ownership or legal status should be checked in the record rather than inferred from this list.

  • US8648867B2 — GPU accelerator, granted 2014.
  • US9189828B2 — GPU accelerator, granted 2015; related patent family.
  • US9626566B2 — Autonomous robotic control, granted 2017.
  • US20170076194A1 — Hardware-agnostic brains for autonomous robots, published application, 2017.
  • US20160198000A1 — Distributed runtime architecture, published application, 2016.
  • US12288162B2 — Deep neural networks with on-device learning, online and offline, with and without supervision, granted 2025.

Questions about Physical AI

Is Physical AI the same as robotics?

Robotics is a major part of it. Physical AI also includes intelligence integrated into instruments, industrial equipment, vehicles, and other systems that sense and affect the physical world. The important question is how perception, decisions, and physical consequences are connected.

Does Physical AI require a large language model?

No. A language model can help interpret an instruction or plan a sequence, but a useful physical system may rely on smaller perception models, state estimation, conventional control, or combinations of these. Its suitability depends on measurable behavior in the intended environment.

Why does simulation not settle the problem?

A simulator is a model of the world. Errors in friction, contact, lighting, timing, or sensing can be enough to break a policy that looked reliable on screen. Sim2Real requires tests on hardware, measurements of the remaining gap, and operating limits.

Does neuromorphic computing mean copying a whole brain?

No. Biological principles can inform a circuit, learning rule, architecture, or sensing method without reproducing an entire nervous system. The engineering case rests on what those choices deliver for the task and its resource budget.

How does my NASA and DARPA work relate to today’s Physical AI?

It connects through a persistent problem: getting learning systems to perceive, adapt, and act under real constraints. The NASA accounts, university reporting, collaborative papers, and patent records above document parts of that history. They do not imply that every subsequent system uses the same architecture or that research intended for planetary navigation flew on a Mars rover.

About the author

I am Massimiliano “Max” Versace, a computational neuroscientist, inventor, and AI entrepreneur. My work spans university research, co-founding and leading an AI company, and leading Emergent AI at a global semiconductor company. The thread running through it is the relationship between intelligence and the physical machinery that makes it possible. For my background, see About Me and my executive profile. For interviews, research discussions, or speaking, contact me.

You missed

Articles Physical AI

Physical AI Explained: What 15 Years of Building Robot Brains Taught Me

November 1, 2026 Max Versace
Physical AI

From Computational Neuroscience to Physical AI

October 17, 2026 Max Versace
Physical AI Talks

Edge AI is not a place, it’s a constraint

July 1, 2026 Max Versace
Physical AI

How AI-powered robotics helped shape NASA’s research and made it into our Earthly devices

April 28, 2026 Max Versace