Animesh Garg joins microagi as Chief Research Officer

The Georgia Tech professor and former NVIDIA and Apptronik researcher is bringing his team to microagi, to build post-training together. His bet: robotics is now a data problem, and that is where it gets solved.

MUNICH, 29 September 2026. Animesh Garg, a pioneer in robot learning, is joining microagi as Chief Research Officer along with the team from Usable Robots.

Garg, whose work has been cited more than 18,000 times, was a Senior Staff Research Scientist at NVIDIA and, until last year, Chief Scientific Officer at humanoid maker Apptronik.

At microagi he will lead post-training: the step that takes a capable general-purpose robot model and turns it into a robot that does one specific job, in one specific place, reliably enough for a company to depend on it. It is the layer the company is betting will decide who gets robots out of pilots and into production, as it pushes toward its stated goal of being the first to deploy a million robots in the real world.

“I’ve been working on the problem of making robots go from the lab to the real world,” Garg said. “We are here to enable thousands and millions of robots to be able to deploy in the real world without the overhead of engineering.”

A shared bet

It started as a collaboration. For eleven weeks, before any terms were discussed, Garg and microagi’s research team worked side by side. By the end, both sides wanted the same thing: robots that leave the lab and stay on the line.

They were already working on the same problem from different ends. Atlas fine-tunes models on customers’ operations and puts robots to work. shift collects first-person video of everyday manual work across 15 countries. Garg’s group has spent years on the research that turns footage like that into robots that keep improving after they are deployed. Doing that as one team, rather than as two companies talking to each other, is the point.

“Frontier AI in robotics is bottlenecked by data much more than it is bottlenecked by compute,” Garg said. “Bringing robots in the real world will require an innate understanding of how to collect good data and at scale. And microagi has shown that it excels at the ability to collect data.”

What post-training is, and why it matters now

Post-training is the least visible and, right now, the most valuable step in getting a robot to work. It helps to separate the two halves of how modern AI gets built.

Pre-training is what the big labs do: feed a model enormous amounts of general data so it ends up broadly capable. It is expensive, it is done a handful of times, and it produces a generalist. Post-training is everything that comes after: taking that general model and teaching it one specific job, in one specific place, until it does it reliably enough that a customer will pay for it and keep paying. In LLMs this second half is what turned an impressive demo into a product people use every day. In robotics it barely exists yet.

That gap is where the commercial value sits. A robot foundation model on its own is a capability. A robot that runs the same assembly step on a specific line, every shift, without an engineer standing next to it, is a contract. Post-training is what turns the first into the second, and it has to be done again for every task, every site and every customer. Whoever does it fastest and cheapest, at scale, owns the deployment.

Garg’s analogy is physical skill. “Think about any skill you can think of, whether it’s playing tennis, or pottery, or painting. We learn to do this with our hands through practice. Without this practice, you do not get to be a master,” he said. “We are betting on post-training for robots with multimodal data in both simulation and real world to achieve this mastery.”

It matters now because three things lined up at the same time. Robot foundation models became good enough to be worth specialising. Hardware got cheap. And it became possible to collect robot training data at scale, which is what microagi built shift for. Until then, post-training was an academic question, because most robots only worked in demos. Garg thinks that has changed.

“We are probably in the most exciting time in robotics as there has been over the last 30 years,” Garg said. “Frontier AI now works. Hardware has gotten cheaper than your computer. And more importantly, we have figured out how to collect robot data at scale.”

The gap that remains is between a good demo and a robot that does the same task every shift on a production line. Closing it is what microagi was founded to do, and it is what Garg has spent most of his career on.

His approach rests on three things. Robots pick up common sense from large volumes of human video, without anyone labelling what the hands are doing. They rehearse in simulation, in closed-loop environments that mirror the real task. And they keep improving after deployment, from their own attempts and from data already collected. “We are betting on recursive improvement that allows every robot to become a specialist,” he said.

Who Animesh Garg is

Garg has worked in robotics for more than 15 years. He did his PhD at UC Berkeley and a postdoc at Stanford, then joined the University of Toronto and the Vector Institute as a professor, holding a CIFAR AI Chair. In parallel he spent six years as a Senior Staff Research Scientist at NVIDIA Research. From 2024 to 2025 he was Chief Scientific Officer at Apptronik, the Texas humanoid company. He is on the faculty at Georgia Tech, where he leads the People, AI and Robotics group.

Some of that work has become infrastructure. He was one of the researchers behind ORBIT, software that lets robots practise a task thousands of times in simulation before touching real hardware. It became Isaac Lab, NVIDIA’s open-source robot learning framework. His 2023 paper ProgPrompt was among the first to show a large language model could write step-by-step plans a robot could carry out.

His more recent work points directly at what he will do at microagi. One line teaches robots from ordinary videos of people working, with no one labelling what the hands are doing. Another has robots improve from data already collected and from their own attempts, instead of needing a person to demonstrate every task again.

“Think about how humans learn. We don’t learn by practicing everything from scratch every time,” Garg said. “Egocentric data is a particularly good mechanism to learn this common sense without specifying each and individual detail.”

What he will build at microagi

microagi already runs two product lines. shift collects data at scale. Atlas gets robots working on customers’ own operations, drawing on microagi’s corpus, and is onboarding large-scale deployments with industrial customers. Garg’s group becomes the third: a research pillar that develops the company’s understanding of pre-training and builds new methods for post-training.

The ambition is stated plainly. “At microagi, we have a very simple goal,” he said. “We want to be the first company that deploys a million robots in the real world.”

Whether that number is reached will depend less on any single breakthrough and more on the work Garg now leads: making each of those robots good at its job, on its shift, in its factory. After eleven weeks of working together, both sides believe the pieces to do it are finally in one place.