Written when I was CEO of Neurala.
At RoboBusiness 2019 I discussed a new paradigm of data collection and training for artificial intelligence.
(originally published here)
Too often, whether it’s an assumption about the power of artificial intelligence, or something seen in a Hollywood movie – roboticists and companies are making mistakes when implementing AI. For example, on the screen, someone downloads AI into a robot, and it wakes up completely autonomous, fully functional, knowing everything needed for a lifetime of service.
Back in the real world, things are more difficult. Companies are spending millions on AI with robotics, only to find that they need a different approach. I talked about a new way to train AI at RoboBusiness 2019, held Oct. 1-3 in Santa Clara, Calif.
Before the talk, Robotics Business Review asked me about it and about the big mistakes robotics companies make with AI.
Training AI differently

“Many companies approach robotics by devoting a ton of time to the body of the robot and the sensor package, and then figuring out the AI at the very end,” I said. “That’s a huge mistake – the biggest misconception that people have is that ‘I can train my AI in the factory and then deploy in the field and it’s going to work.’ That’s the paradigm that everybody’s following, and everybody is failing because in reality, that’s not how you feed the AI.”
Roboticists often underestimate the amount of time and money it takes to develop AI, but they also overestimate the intelligence of what the brain can do, I said. “So they believe that AI is a side dish, with the main portion being hardware, because that’s what they know and what they deal with on a daily basis,” I added. “That’s probably the biggest mistake, the warning that they need is that they need to consider AI at the source from the very beginning.”

A second mistake that robot companies and roboticists make is thinking that once an AI is trained and deployed, it will just continue to work until the end of time. I said that AI is designed to be like the human brain, to continuously learn. Instead of collecting a huge amount of data in the beginning, I said companies need to create initial AI models and infrastructure so it can evolve and refine its knowledge over time.
Continuous learning approach
“It’s like you’re having your last supper – the last meal that you’re ever going to eat – you’re going to eat the chicken with the whale and then the cow – you’re not going to eat forever,” I said. “When you change the paradigm, and have a meal three times per day – at that point you’re not going to have smaller portions. That’s the idea of continuous learning – where you can have the ability to continue to seamlessly update. You don’t need to ask for 7.3 trillion data points at the beginning, because tomorrow you will have 1.7 million more data points, and you can increase the procedure. You and I don’t learn everything at once – learning is distributed throughout the existence of the AI system, not in a one-time fee.”
A third mistake is the belief that AI will make robots smarter, more autonomous, and completely replacing humans in their work. The goal should be for robots to work with humans, taking on some tasks to make things more efficient.
“It’s not going to substitute 100% of the human efficiency, it’s going to supplement the human work for X number of hours, or take a little piece of the work that the human does,” I said. “[The assumption that robots] would substitute the complete human would be premature for where the industry is, and creates false expectations from the customer and end user that the robot is actually equivalent to a human, and we’re not there as an industry.
I gave an example of how AI can be helpful when added to a robot. Neurala has worked with Badger Technologies and its Marty robot, which scans grocery aisles looking for spills, among other tasks. Prior to the introduction of AI, the company would have a team of people looking at the media stream from the robot to check to see if there was a spill or not. “What the AI has done is reduced the amount of time that humans have to look at the video,” I said. Instead, the AI presents the team with images where there’s a high probability of a spill, and the humans don’t have to sift through images of very clean aisles. “The synergy works where AI complements the human takes a part of the job that is dull and just not interesting.”
For my RoboBusiness talk, I said I hope that attendees will walk away thinking of a data as a continuous acquisition and refinement of the AI model, rather than a “last supper” that you can only collect at the beginning and not eat any more. “The software industry has moved to continuous iteration of software a long time ago,” I said at the time. “AI should also move into the continuous collection and iteration improvement the same way.”
I keep everything on Physical AI in one place, from papers and patents to 15 years of press: Physical AI.
More on this: How to train your Robot’s AI · Physical AI Explained: What 15 Years of Building Robot Brains Taught Me
About this recording. “Max Versace from Neurala speaks at RoboBusiness 2019”, 19 minutes, published on YouTube on 8 October 2019. Watch it on YouTube.
Edited and condensed transcript
Cleaned up from the automatic captions and shortened. Cuts are marked […], and where the recording is unclear the wording is my best reconstruction. For the exact words, watch the video.
Thank you for the invitation to speak at RoboBusiness. Today we are going to talk about changing the paradigm of how we train robots.
A little about myself. I do neural modeling, so I am not the other Versace, who does another kind of modeling. I was born and raised in Monfalcone, a coastal town in the north of Italy. I ended up studying artificial intelligence, with a couple of PhDs, one in Italy and one at Boston University, where I founded the Neuromorphics Lab with two colleagues. […] The goal was to build brain-inspired algorithms that learn in real time, different from traditional AI in that they are inspired by the way the brain works. Out of this research I co-founded Neurala. […] So the topic is: what did we learn from the successful applications, some of them in robotics and drones, and what did we learn from the unsuccessful ones?
I live in Boston and I came here to the Bay Area. These are two of the places with the most robotics research, and during the trip I did not see any robots doing anything practical. There is a gap between what we see on YouTube and our day-to-day reality. Many people spend time on issues such as Asimov’s laws of robotics, but that is not the problem we face today. […] Today’s trouble in robotics is that robots don’t work. It is useless to focus on robot laws if the robots we build do not really work.
There is a mismatch between what people expect from robots, which is to be as smart, useful and dexterous as humans, or at least as animals, and the reality, which is that robots are brittle and rigid. […] I will talk about three issues that keep us from getting AI into robots.
The first: “AI is easy.” I like to think of a robotic application as solving three classes of problems at once. I call them the body, the brain and the mind. The body is the chassis, the actuators, the sensors. The brain is the right computing hardware to process information and move around. The mind is the algorithm, the software that powers it all. The usual attitude in the industry, unfortunately, is to leave the mind for last: spend a ton of time and money designing the body, and then we will see about the mind.
We have seen this trajectory over and over. First, time and money go into the body, and that is logical, because a robotics company is there to build a physical machine. Then they realize they need the AI part. And here is the most dangerous step: “we will do it ourselves, it is just AI.” Somebody gets crowned the AI guru. Six months later he is stuck and says: give me another week. Many weeks later the robot still does not work. […]
The second: never underestimate the brain. The brain is about 2% of the body weight and uses about 20% of its energy. Something is going on there. The main thing the brain does is not only understanding the world in front of it. It is learning, constantly. We do not learn once and then stop. We learn in every second of our life. The image I flashed at the beginning of the talk: you remember it, and you remember it because you learned it. You did not just look at it. You wrote it into your synapses.
With my CTO we went through the exercise of calculating how many times a brain learns in a lifetime. Take about a hundred billion neurons and five hundred trillion synapses, about two billion seconds of life, and one spike per neuron per second, which is a conservative number. You get about one septillion learning events, which seems to be about the number of stars in the universe. That is a staggering amount of learning, and it is why the brain uses so much power.
What do we do with robots? We don’t do that. We realized it working with NASA many years ago, when they asked us to design an AI that would let a robot such as a Mars rover explore on its own and learn in real time in a completely foreign environment. The AI available then, which is also the main AI built today, is rigid: you collect a large set of images, you train until the error is satisfactory, and you say it is ready. If you need to add one new thing, you add the data and retrain on the whole set.
We had to come up with a different paradigm, closer to how the brain works: continuous learning. When you see a dog or a chair, you understand it and you learn it at the same time. We call our technology Lifelong-DNN. It learns very quickly and it can add knowledge on the fly, on something as small as a cell phone.
Today, when we train a robot, we collect a huge amount of data, we shower the robot with it all at once, and the robot is done. Compare that with continuous integration, which is everywhere in the software industry: small incremental changes, deploy, test, roll back if there is a problem. That way of working has not made it into AI yet. When it does, robotics will benefit the most.
The third: the right balance between robots and humans. Think of a robot working together with a person on a task, not of a robot that does everything.
One example is a customer whose robots roam grocery stores and inspect them all the time. One thing we did with them is let the robot detect spills on the floor. Before, a human analyst had to look at millions of images per day. Our AI pre-screens them and shows the analyst the small subset that most likely contains a spill. Look at the whole system: the worker in the store, the robot, the analyst. The AI sits in the middle. The result is that the worker in the store has more time for what they are supposed to do, which is customer service, and spills get cleaned up faster.
Drone inspection is similar. The drone collects the data on wind turbines, pipelines and power lines, and the AI does the tedious part, so the analyst looks only at the images that matter. Another customer builds robots that go underground, for example in a disaster area, where you do not want a person to go first, to see what is there and whether there are survivors.
And then manufacturing. People do not think of industrial machines as robots, but they are becoming more and more robotic. […] A machine puts out sixty products per minute, around the clock. A person cannot keep up with that for quality control. With AI doing the fast, tedious part, one worker can oversee several machines, in much better working conditions, and look only at the products the AI flags.
To summarize, three keys. First, plan the AI ahead. It is not an easy task: build a large team, or work with a company that does this, but do not leave it to the end. Second, move away from training your AI once and hoping for the best. AI is software that has to be updated, and your roadmap should include that. Third, and this you should think about before anything else: what is the real role of the robot, and of the humans, in the task? Design the robot not to do everything, but to do enough that the human is relieved of a crucial part of the work.
Thank you again for the invitation.
New here? Start with what Physical AI is, the glossary or the starter kit.

