In a robotics lab just north of Lake Union, a robot arm hovers over a tabletop. It has just stacked a wooden block marked “i” onto one marked “2,” leaving it slightly off center. Four roboticists look on as the arm’s two-fingered metal gripper fumbles for the last block, an “A,” to spell the name of their nonprofit institute — Ai2.
A preschooler could do it. Anyone who has watched robots fold laundry or do kung fu in online videos might wonder how this counts as progress.
But the work happening in the lab could dramatically change how robots powered by artificial intelligence learn. Most robots are trained today by people steering them through a task again and again, which is slow and costly. Getting 10 times more examples means paying for far more hours of human labor.
“We don’t think it’s scalable,” Florian Shkurti, an Ai2 team member and visiting research scientist from the University of Toronto, said of this traditional approach. “And we don't think it's the way academia and research should go about this.”
Ai2’s solution is ToyBlocks, which researchers describe as a “virtual lab.” One thing ToyBlocks does is create computer simulations of the real lab setup Ai2 has in Seattle — the table, the blocks and the arm.
Instead of a person guiding the arm through every stacking and flipping movement, the software underlying ToyBlocks automatically generates thousands of different practice runs. The AI model that controls the robot learns from these computer simulations. Researchers can then test how those skills carry over to the real arm.
On Thursday, Ai2 released ToyBlocks for free public use. This means anyone can train and test robots on the identical setup and build off it.
Until recently, many roboticists doubted that simulation data was high quality or abundant enough for this approach to be effective.
That changed this year, said Russ Tedrake, a robotics professor at the Massachusetts Institute of Technology, and co-founder and CEO of Walden Robotics. Powerful, new AI models from OpenAI and Anthropic made it possible to build detailed simulated scenes on request.
Longtime skeptics are now suddenly hiring simulation engineers," he said.
Robotics development has long been slowed by real-world experiments, Tedrake explained. If it were to only be limited by data and computing power, he said, “you can just let it rip.”
Ai2, long known for its open artificial intelligence research, is newer in robotics. Last year, it hired Dieter Fox from Nvidia's Seattle robotics lab, and expectations were high. With a team of about 15 that started as just Fox and two others, Ai2’s strategy was to establish ToyBlocks for others to use and build on, to more systematically advance the robotics field.
"It's not just us doing this; we are way too small for that," Fox said. "We need the help from the scientific community."
Getting robots to learn on their own in a virtual setting is challenging.
Ai2 researchers gave a leading robot AI model, from San Francisco startup Physical Intelligence, 1,000 simulated examples of stacking a green block on a red one. Tested on the same patch of table, it succeeded at stacking almost every time.
Asked to stack the blocks on another part of the table, the robot pretty much failed.
"For a person, this would be trivial," Fox said. But the robot hadn't learned to stack blocks. It had learned to stack blocks in one spot.
ToyBlocks is a place for researchers to test whether more data fixes this kind of problem. “What happens if I gave you a million examples?” Fox said. But, he added, quantity isn’t everything: “What really matters is the right kind of diversity in that data.”
Eventually, Fox imagines a robot being let loose in a virtual world for a month just to watch what it teaches itself, the way children learn by playing with blocks.
An added benefit with ToyBlocks is that it offers a standardized benchmark to compare robots against one another. It allows researchers to change one thing at a time, like where the blocks sit or how big they are, to pinpoint precisely where a robot fails in completing a task.
For roboticists, this evaluation framework is an important development.
“There’s no real way to say, ‘Are my robots better than your robots?’ " said Tedrake.
Because of recent breakthroughs with artificial intelligence, some robotics researchers are hoping that the simulation training might lead to a ChatGPT-style leap for robotics.
In the meantime, Fox said, the goal is simply to build better foundational tools for robotics research to put it “on a more scientific footing.”
While he described progress with simulation-guided training for robots as "quite dramatic," Fox wouldn’t predict when robots "people often dream about will arrive — machines that can learn on their own to navigate dynamic, changing situations.
But with goals like that in mind, the research underway at Ai2 in Seattle is about much more than stacking blocks.
© 2026 The Seattle Times. Visit www.seattletimes.com. Distributed by Tribune Content Agency, LLC.
