Robotics startups get the most from an academic hire when the professor owns one commercial problem and arrives with the researchers who already make their lab work. Microagi did both in September, naming Georgia Tech professor Animesh Garg chief research officer and bringing seven of his researchers in through an acquisition.
Garg will run post-training at the German startup, the stage where a general-purpose robot model learns the specific tasks on one customer's line. Microagi builds those models from data captured by factory workers wearing cameras and sensor gloves, and it closed a $55 million seed round this summer. That's the work that decides whether each new customer goes live, and giving it to a professor changes who signs off on a launch.
Physical AI has reached the point where a robot model trained in a lab can earn revenue on a factory floor, and a large share of the people who know how to get it there came up through university labs. Garg described the shift as a threshold crossed, with the technology now able to support sustainable businesses.
Skild AI, cofounded by Carnegie Mellon professors Deepak Pathak and Abhinav Gupta, has raised more than $2 billion for a single model that can run many kinds of robots, while Physical Intelligence, cofounded by Stanford's Chelsea Finn and UC Berkeley's Sergey Levine, has passed $1 billion. This summer MIT professor Russ Tedrake brought Walden Robotics out of stealth too.
A Series A robotics company hunting for someone who has trained manipulation policies on real hardware ends up chasing the same postdocs and senior PhD students as those founders, while the research groups at the big AI labs want the same people. For startups without a well-known research founder, a stronger pitch can be a specific deployment problem with real customer data behind it, plus a seat reporting directly to the CEO.
The role becomes necessary once the biggest unknown in the business is a model problem. For Microagi, that's whether a general model can be adapted to the next customer's line fast enough to sell, so post-training went to the chief research officer. Keeping robots running on a customer site stays with engineering.
Until then, a research founder in the CTO seat is common and usually fine, at least while the company has one prototype and a small engineering team. The strain shows up once customer deployments and model work are fighting over the same person's week.
In practice, the head of research decides which models to build and what data to collect from workers on the line. The CTO owns fleet uptime and the field engineers who install robots on customer sites.
Where the two roles run into each other is evaluation: if research alone calls a model ready, field engineers end up holding the failures halfway through a customer's night shift. A CTO who keeps overruling launches causes a different problem, because research stops hearing about what broke. Most of that fight goes away when both sides agree before the hire that research sets the bar and engineering confirms it on site.
Garg's researchers had been building a startup in stealth, so the acquisition brought in a team that already shares code and working habits. Hiring eight people one at a time would take months to reach the same rhythm.
If the group reviews only its own code and still answers to its former advisor, though, it can turn into a separate company inside the company, with product engineers hearing about its work at demos. Students and postdocs also arrive with academic expectations about publishing and authorship, and those belong in the deal terms.
Before signing, a CEO and CTO should have answers to these:
An advisory or part-time arrangement, with the professor keeping a faculty post, is a cheaper way for earlier-stage companies to test the fit. Advisors rarely own outcomes, though. If what's missing is a research leader, an advisor won't fill the seat.
A strong publication record shows someone can move a field forward, though it says little about getting a model working in a food plant by the date a customer was promised. Garg is an easier case, after six years as a senior staff research scientist at Nvidia, where his simulation work ended up in the Isaac Lab training platform. From 2024 to 2025 he was chief scientific officer at Apptronik, a humanoid robot maker.
Candidates without that kind of industry time should expect harder questions:
Professors already recruit and fund their own groups, and that's most of a research leader's week. Pace is the harder adjustment for someone used to grant cycles measured in years who now answers to customers expecting results every quarter.
For references, former students now in industry know how the candidate behaves near a deadline. So does any company that sponsored past lab work, and it can also say whether the group delivered what it promised.
The first question is whether the professor resigns or takes leave. Leave gives the professor a way back to campus, which can make yes an easier answer. It also comes with a clock and with conflict-of-interest rules the university writes. The startup's counsel should read the candidate's faculty policy before any offer goes out.
Universities often own what gets invented in their labs. That gets complicated when a new hire arrives carrying code, or an entire stealth startup, built there. A written assignment from the university costs far less than a dispute after the first large customer signs.
For some researchers, the right to keep publishing weighs as much as cash, while a company whose edge sits in its post-training methods has good reason to hold some of that work back. The common compromise is a fixed review window before submission, plus a short list of topics that stay internal.
Big AI labs pay senior researchers more cash than a startup can, and equity has to close that gap. With an acquired group, every member should vest against a milestone the team cares about. A first deployment at a new customer is an obvious one.
Chief Research Officer, the title Microagi chose, can signal that research sits at the executive level. VP of Research, Head of AI Research, or Director of Research often means the role reports into the CTO, which suits companies where research mainly serves the current product.
The clearest one is a research roadmap measured in papers while customer go-lives wait on model work. When research's quarterly update lists submissions and benchmark scores with no line about which deployments got faster, the role has drifted back toward academia.
Field engineers usually see the trouble first. If they keep rewriting research code from scratch before it can run on a customer site, the two teams never agreed on what finished code looks like, and evaluation tests written by both sides are a good place to start that conversation.
An acquired group still running its own standups and its own repository six months in hasn't joined the company yet.
A CEO who ends up refereeing every launch between the research leader and the CTO skipped the conversation about who decides readiness. Putting that agreement in writing before the hire starts gives both of them something to revise after the first failed deployment.
The role owns the direction of the company's models and the team that builds them. That can include post-training, where general-purpose robot models are adapted to the tasks on a specific customer's line. At some robotics startups, a chief research officer reports to the CEO and works alongside the CTO.
It can work at the prototype stage, when the main problem is research. As fleet operations and customer integration grow, companies may separate the roles, especially when research and deployment begin competing for the same leader's time. A professor who has shipped robots in industry is the stronger CTO candidate.
Sometimes, through a leave of absence or a reduced appointment. Each university sets its own limits on outside time and conflicts of interest, so the rules are worth reading before negotiations start.
The right moment is when the model itself is what's holding back sales. That often arrives after the first paying customer, once winning the second depends on adapting the model faster.
