Why 2025 is a pivotal year for robotics

It’s cute, small, goofy-looking and programmable in Python. Meet Reachy Mini, a compact, open-source robot from Hugging Face designed to be used by developers, regular consumers and even kids. And since Hugging Face announced the release in July, Reachy Mini, available on pre-sale, has exceeded expectations, according to Pollen Robotics, a robotics company recently acquired by the AI company.

2025 is shaping up to be the year of physical AI. According to NASDAQ, the market is projected to grow from USD 12.77 billion in 2023 to USD 124.77 billion by 2030. NVIDIA also announced the first open humanoid foundation model, GROOT N1. And robots are already being put to work: Amazon just deployed its millionth robot in July 2025, and Google’s Waymo delivered over 10 million rides by May 2025.

But are robots ready for everyday life? IBM Think met with the founders building toward that future.

Anto Patrex left his job at a major AI company this year to launch CosmicBrain, a startup tackling one of robotics’ biggest challenges: data.

“A lot of people started building and using AI to enhance robotics,” he said in an interview with IBM Think. “So that’s one of the reasons why I quit my job. I felt if I’m not in the space as of today, I’ll be missing out on a lot.”

CosmicBrain builds simulation models to train robots using human task data, collected via any wearable smart glasses. “If you look into robots from physical intelligence companies, they are getting trained using teleoperation,” Patrex explained. “This is not a scalable method, so when they are deployed into an actual warehouse or agriculture field, it doesn’t have a diverse dataset.”

The idea? A plug-and-play app store for robots, for everyone: developers, students or companies willing to train their robots. “We are creating a Google Play Store or an App Store for robots,” he said. “It’s like the modern-day Matrix. We’ll have hundreds of thousands of skillsets you can just plug in, and the robot will know what to do.”

Patrex said the plan is to collect over 50 million hours of video in the next two years. The next step will be to use this data to generate synthetic data.

“There’s a lot of stuff that we can’t actually do in the real world,” he said. “What happens if the robot lets a package slip down? How does he pick it up? We can teach these robots to do this, or to recognize 100 varieties of different fruits, like picking up cherries versus tomatoes, apples and oranges.”

CosmicBrain sees its potential impact on the robotics ecosystem similar to the one Scale AI had, which specialized in data annotation, and  Google, Microsoft, Meta, OpenAI or Anthropic had on the generative AI boom. “Except Scale AI sold their data,” he said. “We will not be selling our data. Rather, we'll be selling skill sets as APIs, so the data will stick with the company.”

A few days ago, Elon Musk, Tesla’s CEO, predicted that Tesla’s Optimus robots could bring in USD 30 trillion in revenue, and Patrex buys into it. “Soon, you will see [everyone] … building robots, repairing robots, painting robots, training robots,” Patrex said. “There’s going to be a huge economy.”

That future is already materializing for Jan Liphardt, an Associate Professor of Bioengineering at Stanford and Founder of OpenMind, an open, AI-native software stack for robots that plans to open its first shop in San Francisco this year.

“This is happening this year, not in 2037,” Liphardt told IBM Think. “And it’s not just us.”

A longtime fan of sci-fi robots, Liphardt observed the gap between fictional robots and the industrial machines in real life. “Many of us have seen movies where people interact with robots,” he said. “And until a few years ago, robots were built to move a piece of metal and weld it to another piece of metal, for industrial purposes. There was this humongous gap between the kind of robots that we saw in movies and read about in science fiction, with the kind of robots that we saw around us.”

That changed during the generative AI boom.

“I was looking at LLMs just like a few years ago,” he said. “I was curious to see if the set of outputs can be expanded from text to actions in the real world.”

He started writing software to control home robots—helping kids with homework, barking at unfamiliar guests, even mapping parks. “LLMs not only can write computer code, but they can also issue commands to hardware,” Liphardt said. “It was really exciting for me.”

He believes LLMs solve core challenges like data fusion, the process of integrating data from multiple sources to produce a cohesive and informative output. “One advantage of using language to describe the physical world is that it fits naturally with how large language models work—they’re great at understanding, continuing and generating stories from that input.”

Still, not everything works as intended. One of his robots, for example, tends to bark at homeless people. “Of course, we didn't program that in,” he said. “But the idea of a homeless person being scary is something that must be in the training data.”

“This [bias in training data] is not specific to robots,” Liphardt said. “And what’s important for me is simply to observe that and then think about good ways of controlling or governing the technology. And that's also a completely unsolved problem. If you think about the intersection of alignment governance and large language models, that’s a much, much bigger problem than any robot software effort.”

Driving developer adoption might be what finally pushes robotics into the mainstream. That’s the bet behind Hugging Face’s move into the space. “With cheaper hardware and open-source AI, we think this could be the ChatGPT moment for robotics,” said Hugging Face CEO Clément Delangue in a recent podcast episode of TBPN.

K-Scale Labs, an open-source humanoid robotics company based in San Francisco, is working on the same premise. They will start shipping K-Bot, a personal robot that will have basic locomotion, voice control and a pre-defined command set, in December 2025.

“I think it's very important to have a really open source developer first platform,” said Benjamin Bolte, the Founder and CEO of K-Scale Labs, in an interview.

A former Tesla and Meta engineer, Bolte believes that, like the early days of PCs or game consoles, the real growth will come from early adopters, developers and hobbyists.

K-Scale Labs is betting on a few key use cases—and on openness. “Trying to dictate how everything should work is the wrong approach,” Bolte said. “If we can actually have people experimenting with robots and actually building the product experience themselves, I think there's going to be like a large downstream, like an app developer ecosystem for humanoids.”

Building on the interest for Reachy 2, a fully open-source, humanoid robot developed and built by Pollen Robotics, the company developed Reachy Mini, a personal robot only available on pre-order. This is not a full humanoid robot, but potentially a day-to-day assistant that can move around your house, talk, listen and visualize its environment.

“The range of applications is huge,” said Anne-Charlotte Passanisi, a Senior Product Manager at Pollen Robotics, the robotics company acquired by Hugging Face earlier this year. “We’re building the blocks to let people create amazing things. The goal is a family-friendly robot anyone can program. We’re not there yet, but the potential is enormous.”