OpenAI Just Made It 7 Buyers for 45 Engineers. Title Search Is Done.

OpenAI's late-May and early-June 2026 robotics postings (perception, controls, simulation, data) ended the polite version of this market. The humanoid talent fight was already crowded with Figure, 1X, Apptronik, Agility, and Boston Dynamics. With OpenAI in (and Google DeepMind quietly hiring through its Apptronik partnership), there are now seven buyers chasing a pool small enough to fit in a single conference room.

If your sourcing plan still starts with "Robotics Engineer" as a LinkedIn title filter, you are searching the wrong index.

The 7-buyer market, in one paragraph

Figure shipped Helix v2 in February and is running paid pilots inside BMW Spartanburg, where it has logged 30,000+ X3s supported in 11 months, 90,000+ sheet-metal parts handled, roughly 1,250 operating hours, and 1.2 million steps. 1X moved from Neo Beta into Neo Gamma consumer pilots in late Q1, with Early Access units shipping to U.S. homes in 2026. Apptronik has Apollo running inside Mercedes-Benz, GXO, and Jabil sites, plus a strategic partnership with Google DeepMind on Gemini Robotics (that is your seventh buyer, hiding in plain sight). Agility is in Amazon fulfillment. Boston Dynamics retired hydraulic Atlas, shipped the electric variant, and is running it in Hyundai plants. OpenAI's new robotics org, currently without the hardware lead it had until March 7 when Caitlin Kalinowski resigned over the Pentagon deal, is hiring into all of the above.

Industry analysts expect 50,000 to 100,000 humanoid units to ship in 2026. The engineers who can actually deliver against that number are not browsing job boards.

The pool is 45 people, and Apple is tied for first

Here is the number that should reframe your entire 2026 sourcing plan.

45

That is the true addressable pool if you source by title intersection. Forty-five people in the entire United States who carry a "Robotics Engineer," "Controls Engineer," or "Robotics Research Scientist" title and have RL on their profile. Concentrated in the Bay Area. Two at Apple. Two at Boston Dynamics. The rest scattered across Agility, Dexterity, Figure, 1X, Apptronik, and a long tail of stealth startups.

Seven buyers. Forty-five people. You can do the math on what happens to comp.

You can also do the math on what happens to your time-to-hire. The Global Robotics Report 2026 puts specialized robotics vacancies at an average of 114 days open. OpenAI will close in days. Everyone else is running quarter-long loops against a candidate who already has three competing offers.

"Robotics Engineer" is four jobs in a trench coat

The KORE1 breakdown is the cleanest framing of why title search fails here. "Robotics Engineer" in 2026 means at least four distinct profiles:

• MuJoCo / Isaac Lab trajectory and controls engineers. Whole-body control, MPC, contact-rich manipulation. These are the people who own the gap between sim and a real arm that does not snap a wrist (Figure's forearm failure mode, fed back into the Figure 03 wrist redesign, is the canonical example).
• ROS2 platform engineers. Middleware, real-time scheduling, fleet ops. Less glamorous, completely necessary, paid like infrastructure.
• VLA-training perception engineers. The Helix profile. Vision-language-action model training, multi-modal data curation, on-robot fine-tuning. The most-poached profile in the market right now.
• MoveIt 2 motion-planning specialists. Classical planning, collision checking, integration with learned policies.

None of these people interview the same way. None of them cost the same. None of them are interchangeable in a JD. If you post one role that mushes all four together, you will sit at 114 days and then hire the wrong one.

This is the specific friction Refolk was built for: you describe the human you actually want ("VLA training experience on a real humanoid, has touched MuJoCo or Isaac Lab, published at CoRL or RSS in the last two years, currently in the Bay Area") and get a ranked shortlist that ignores title noise. Title-based LinkedIn sourcing is structurally broken for this market because the title compresses four different jobs into one string.

Where the real signal lives

If the title index is broken, what is the index?

GitHub commits to the right repos

Isaac Lab, MuJoCo, LeRobot, Drake, Pinocchio, and the various Isaac Sim extensions. The contributor graphs of these repos are a better proxy for "can ship sim-to-real" than any LinkedIn headline. A senior IC who shows up three times in the Isaac Lab commit log is more interesting than a director whose only signal is a title change.

Paper authorship at CoRL, RSS, ICRA, and CoRL workshops

The pool of people who have first-authored a sim-to-real paper at CoRL in the last 24 months is smaller than the pool of people who say they do robotics on LinkedIn, by something like two orders of magnitude. It is also the pool that Figure, OpenAI, and DeepMind are recruiting against.

Discord and HuggingFace community signal

The LeRobot community on HuggingFace, the Physical Intelligence (π) Discord, and the active threads around Isaac Lab releases are where the under-30 specialists actually post. None of that surfaces in a Recruiter search.

OpenAI's entry is a speed problem, not a salary problem

The reflexive read on a new buyer is "comp goes up." That has already happened. Robotics salaries are up 25 to 40% since 2023. The 2022 candidate who said yes to $180K base now has a competing offer at $260K base plus pre-IPO equity, and the recruiter on the other side has been working him since SXSW.

The actually-binding constraint is closing speed. OpenAI's brand pull means it can move a senior IC from first conversation to signed offer in under two weeks. If you are a Series B humanoid lab running a six-stage loop with two onsite days and a take-home, you will lose every contested candidate to OpenAI on timing, not money. The differentiator for non-OpenAI buyers in this cycle is interview architecture, not salary bands.

The differentiator for non-OpenAI buyers is interview architecture, not salary bands. OpenAI closes in days. You have weeks.

Concretely: collapse the loop to three stages, pre-commit your comp band, and have the founder or CTO available for a same-week dinner. If you cannot do that, do not start the conversation.

The Figure / OpenAI split is now a recruiting weapon

The Figure and OpenAI partnership ended in February 2025 when Figure announced Helix and took the AI stack in-house. Eighteen months later, that breakup is a sourcing map.

The Figure engineers who built Helix in-house, against a real BMW deployment with 1.2 million logged steps of training data, are the exact profile OpenAI's new robotics org wants. Expect targeted poaching of Figure's VLA team. Expect counter-poaching by Figure of OpenAI foundation-model researchers who want to touch hardware. Expect both sides to deny it is happening.

The recruiters who win here are the ones who can identify the Helix contributors by name (commit log, internal demo credits, conference talks) without waiting for them to update a title. That is the second place Refolk earns its keep: you ask for "engineers who plausibly worked on Helix v1 or v2 at Figure based on talks, papers, or commit activity" and get a list, not a tab full of profiles that all say "Member of Technical Staff."

Don't forget the seventh buyer

Google DeepMind is hiring the same pool, through two channels most recruiters are not counting:

1. The Apptronik / Gemini Robotics partnership, which is a hiring funnel as much as a product partnership.
2. The ex-Boston Dynamics senior-IC pipeline, most visibly Aaron Saunders (former CTO at Boston Dynamics, now at DeepMind). Where the CTO goes, the senior controls people follow within 18 months.

If your competitive map says "5 humanoid labs plus OpenAI," you are mis-sizing the market by one entire buyer. The actual number is seven, and DeepMind is closing offers from inside a partnership you cannot see on LinkedIn.

A 2026 sourcing playbook for the next 90 days

If you are a non-OpenAI buyer in this market, here is what works between now and Q4:

1. Throw out your title filters

Source on skills and artifacts: MuJoCo, Isaac Lab, LeRobot, Drake, MoveIt 2, RL on contact-rich manipulation, VLA fine-tuning, ROS2 in production. Search for the work, not the headline. This is where humanoid robotics recruiting in 2026 actually diverges from 2022: the title index is noise, the artifact index is signal.

2. Split the JD into four

Stop posting "Senior Robotics Engineer." Post a controls role, a sim-to-real role, a VLA training role, and a motion-planning role. Each one will attract a different shortlist, and each one will close at a different comp band. The companies still posting one mega-role are the ones sitting at 114 days.

2b. Mine the conferences, not the job boards

CoRL 2025 accepted papers, RSS 2025 program, ICRA 2026 author list. Cross-reference with current employer. That is your VLA model engineers shortlist for the next two quarters. For sim-to-real engineer sourcing specifically, the Isaac Lab and MuJoCo contributor graphs are more valuable than any sourcing seat you already pay for.

3. Compress your loop

Three stages. Pre-committed comp. Founder available within a week. If you cannot promise a verbal in 14 days from first contact, you will lose to OpenAI on every contested req.

4. Build a watchlist for the seventh buyer

Track Apptronik / DeepMind co-authored papers, Gemini Robotics blog posts, and ex-Boston Dynamics moves. The senior-IC migration pattern is visible six months before the headcount shows up.

The labs that win 2026 are not the ones with the biggest comp bands. They are the ones who figured out that title-based sourcing died the day OpenAI re-entered the market, and rebuilt their pipeline around artifacts and authorship instead. That is the bet behind Refolk: in a market where the addressable pool is 45 people and seven buyers are fighting for them, the team that asks the right plain-English question gets to the candidate first.

FAQ

Is the 45-person pool really the whole market?

It is the pool you can find by title intersection (robotics title plus RL as a listed skill) in the U.S. The full addressable market is larger, but only if you source on artifacts (GitHub, papers, conference talks) instead of titles. That is the entire point: the people you actually want are not discoverable through standard LinkedIn filters, which is why title-based sourcing is producing 114-day vacancies across the sector.

How do I source Figure's Helix team without burning the relationship?

Carefully, and not through a generic InMail. The Helix contributors are identifiable through Figure's own technical posts, BMW pilot writeups, and CoRL/RSS authorship. Reach out with specific reference to their published work, offer a concrete technical conversation with your founder or research lead, and respect that most of them already have three open conversations. The ones who reply are the ones who want to touch a different hardware stack or a different scale of deployment.

Does OpenAI's entry actually change comp, or just close speed?

Both, but the closing-speed delta is the one most recruiters are underestimating. Comp was already up 25 to 40% since 2023. What changes with OpenAI in the market is that a senior IC can now go from first conversation to signed offer in under two weeks. If your loop takes six weeks, you are not in the fight, regardless of your band.

Who is the seventh buyer, and why does it matter?

Google DeepMind, hiring through the Apptronik / Gemini Robotics partnership and through the ex-Boston Dynamics senior-IC pipeline (Aaron Saunders being the canonical example). It matters because recruiters who model this as a six-buyer market will miss where the most senior controls people are actually landing. DeepMind is closing offers from inside a partnership that does not show up as a competing employer on any standard sourcing tool.