Humanoid robots are moving beyond eye‑catching demos and into real-world operations. This shift is being driven less by hardware and more by advances in AI. Speaking at this week’s Humanoid Robot Forum held during A3’s Automate 2026, Aya Durbin, Director of Product at Boston Dynamics, highlighted why general‑purpose humanoids may finally tackle the messy, high‑variability tasks traditional automation can’t.
“Today, I want to talk about what's changed on the technology front that has actually enabled us to go from purpose-built to general purpose,” said Durbin, sharing lessons from observing factory floors.
According to Durbin, there are tasks that have been automated already today in manufacturing facilities with hand-programmed original approaches, but there are also enormous number of tasks in these facilities that we haven’t been able to automate.
“Humanoids aren't meant to take over all work that's already automated. Humanoids are meant to take over work where you need flexible solutions. They can work in the same places and face the same challenges that humans do,” she said. “They can actually do the hard work that exists in industrial environments — the work that’s really hard to hire for and even harder to retain labor for.”
“If you think about the types of tasks in your facility that you haven't able to automate yet are in environments with extreme variance. You need to work with all different types of equipment. You need to move through large areas of facilities that require high dexterity or you work in an environment that's not easy to predict or control,” she said.
What's changed, said Durbin, is that for the first time ever, the technology actually exists to solve these problems, opening the door to more flexible, scalable robotics in manufacturing and logistics.
The shift from programming to training behaviors that has happened is transformative. “You've heard repeatedly that AI is going to solve all our problems and it really is changing what we're able to do with this technology. And I'm here to tell you, it's not a lie. Physical AI allows us to train the robot on a new set of behaviors and get to know what to be reliable at performing those behaviors at a rate that we have never seen before,” she said.
The promise of physical AI for robotics is that you'll be able to do all sorts of different tasks, re-iterated Durbin. “The way that we train the robot to do these different tasks is very similar to the way we train LLM models. It gets better as it gets more data, and it gets better as you use it. AI for robotics works similarly, but it's a little bit more complicated because we're outputting physical behavior as opposed to language, which means our inputs are a little bit different. But the idea is the same. The more data you have, the more skills you teach the robots, the more skills the robot has, the more reliable a robot is at performing that behavior.”
Durbin also discussed how Boston Dynamics uses reinforcement learning — a form of training that's done in simulation — to train its humanoid Atlas. “With this type of training, you show the robot how to do something in simulation, you run 1000s or million of times in simulation, and it results in autonomous behavior.”
“The cool thing about this,” she added, “is if you have a really good sense of real capability, you can disturb the environment.” This means exceptions can be created with scenarios for the robot and teach it how to react in different environmental conditions.
Boston Dynamics also uses “egocentric" video training to train Atlas. “This allows us to take how you do work in your facility today and train the robot directly using just the people and technology that already exists in the buildings. For example, with a GoPro style camera on your head, and either gloves or no gloves on your hands, you can train the robot in different, usually more dexterous-style tasks. What we train in robots becomes what our capability set is,” she explained.
Humanoid robots aren’t immune to the challenges of industrial automation. “As we scale, specifically in the industrial market, we get a lot of benefits, specifically because we're using AI to train the robot. Not only do you get to take advantage of economies of scale, but you also actually get a robot that can take on increasingly complex tasks over time,” said Durbin.
To scale, however, humanoids still need to integrate into industrial systems and that’s equally challenging. According to Durbin, AI might help us solve this problem too: “Imagine an AI agent that would do the system integration for you or automated tools that would actually do the testing for us between the systems so that we can make sure that the robot is going to be performing before it’s even on site. All these parts of the process that are painful can all become significantly more simple with the advent of AI.”
Durbin showed videos of how Boston Dynamics is implementing these solutions and using AI to train its robots and solve problems. “And that's the reason why I've been convinced that general purpose might actually be the more pragmatic solution,” concluded Durbin. Automate Booth #225
For more information, visit www.bostondynamics.com.
