"Robot Learning Engineer" Postings Up 340%. The Real Pool Is 41.

"Robot Learning Engineer" postings have grown 340% since 2024, the fastest of any robotics specialization, and Physical Intelligence, Skild AI, Figure, Agility, and Google DeepMind Robotics are all writing more or less the same job description. The problem: almost nobody who can actually do the work calls themselves that on LinkedIn.

The 102x gap between title and skill supply

Only 41 US profiles currently title themselves "Robot Learning Engineer" or a close variant, while 4,188 US Research/ML/Robotics engineers hold the underlying skill combo of Reinforcement Learning plus ROS plus PyTorch. That is a roughly 102x gap between the title recruiters are searching and the pool that can actually train a Vision-Language-Action model.

A Vision-Language-Action (VLA) model is a neural network that maps camera pixels and a language instruction directly to robot actions, the architecture behind Physical Intelligence's π0 and Stanford's OpenVLA. Sim-to-real is the practice of training a policy in simulation (Isaac Sim, MuJoCo) and transferring it to physical hardware without collapse.

102x

Any recruiter running a Boolean search for the job title is sampling the least representative 1% of qualified candidates. The mechanism is simple: the field renamed itself faster than candidates updated their profiles. Most VLA practitioners were "PhD student" or "Research Engineer" 18 months ago. They have not rebranded, and the ones who did tend to work at the wrong companies.

Where the 41 self-titled actually work (hint: not Physical Intelligence)

The 41 people who self-title as Robot Learning Engineer cluster at Ford, Aramco, Astrobotic, Microsoft, and Huawei, not the frontier VLA labs everyone is racing to poach from. This is the strongest evidence that title-based robot learning engineer sourcing is broken.

Frontier labs still use "Research Scientist" or "Member of Technical Staff" internally, because that is the title that attracts the PhDs they want. Legacy R&D shops adopted the trendy label a beat later because it helps req approval, not because it maps to their internal ladders. If a candidate's current title is literally "Robot Learning Engineer," treat it as a mild negative signal for frontier fit and look at their arXiv output first.

The skill-defined pool clusters very differently. In Refolk's index, the top current employers among Senior+ profiles with RL plus Robotics skills include:

• Meta (6 profiles in the sampled slice, FAIR-adjacent)
• Sesame
• Cursor
• Apple (robotics group)

Those are LLM and foundation-model shops. That is the real poaching battleground for humanoid robotics recruiting in 2026, because a strong VLM researcher is one recruiter call away from a VLA role.

The full Refolk index numbers

Here is the entire dataset, so you can size any sourcing plan against it.

The salary number matters for compensation calibration. Median US robotics engineer salary hit $148,000 in early 2026, a 14% jump over 2024 and 68% over 2020. Frontier VLA offers run well past that, but the median is where competing counter-offers from Ford or Astrobotic will land, and those are the offers you actually have to beat when a candidate is deciding whether to move.

VLA and sim-to-real are two candidate profiles, not one

Job descriptions treat "VLA and sim-to-real" as one bullet point. They are two distinct candidate archetypes, and sourcing them from a single search is why so many robotics reqs sit open for a quarter.

• Sim-to-real practitioners come from RL and control. Tools: Isaac Sim, MuJoCo, domain randomization, PPO, model-predictive control.
• VLA practitioners come from VLM and multimodal LLM training. Tools: PyTorch, JAX, large-scale data pipelines, flow matching, diffusion policies. Papers cite π0, OpenVLA, GR00T.

These profiles rarely coexist in one head. If your JD lists both, you need two sourcing playbooks and two shortlists, not one. Describing that split in plain English is the exact gap Refolk closes: you can ask for "sim-to-real RL engineers with Isaac Sim commits" and, separately, "VLA researchers who co-authored on arXiv in the last 12 months" and get two clean ranked lists instead of a keyword-fried mess.

The arXiv co-authorship shortcut

The pool that has actually trained a production VLA is small enough to enumerate by hand from paper author lists. This is the single highest-signal filter for VLA engineer hiring in 2026.

Scrape authors from these papers, then dedupe:

1. π0.5 (arXiv:2504.16054), Physical Intelligence
2. OpenVLA (arXiv:2406.09246), Stanford, 7-billion parameter open baseline
3. MolmoAct (arXiv:2508.07917)
4. CogACT (arXiv:2411.19650)
5. SmolVLA (arXiv:2506.01844)
6. NVIDIA Isaac GR00T N1.x papers and tech reports

Author lists on those preprints are effectively pre-qualified candidate slates. Cross-reference each name against GitHub for commits to Physical-Intelligence/openpi, allenzren/open-pi-zero, or NVIDIA/Isaac-GR00T, and you have separated hobbyists from people who actually shipped training code.

An arXiv author list on a 2025 VLA paper is a better sourcing surface than any LinkedIn title filter you will ever build.

The π0 training run is a useful yardstick for what "real" experience means. It was pre-trained on roughly 10,000 hours of robot data across 7 robot configurations and 68 tasks, plus open datasets like OXE. Candidates who talk fluently about data curation at that scale have actually done the work. Candidates who talk only about model architecture probably read the paper.

Pittsburgh is the density trade you are missing

Pittsburgh holds 3 of the 41 self-titled Robot Learning Engineers, roughly 7% of the cohort against Pittsburgh's ~1% share of US tech employment. That is the densest per-capita concentration in the country, and it is entirely a Carnegie Mellon spillover effect.

The geographic breakdown of the 41 self-titled cohort:

• San Francisco Bay Area: 4
• Pittsburgh (CMU orbit): 3
• Houston: 2
• Remainder: distributed thin

CMU's Robotics Institute has been shipping formally-trained robot learning PhDs for a decade, and their alumni network holds together tightly through spinouts like Astrobotic and Skild AI. If your sim-to-real recruiter workflow starts and ends in the Bay Area, you are leaving the highest-density source city on the table. A useful filter for physical AI talent pool sourcing: "CMU RI PhD, graduated 2020 to 2024, currently outside Pittsburgh."

The GitHub and HN signals worth actually watching

Three sourcing surfaces produce higher signal than LinkedIn title search for this specialty: GitHub contributor graphs on the canonical VLA repos, arXiv author lists on the canonical papers, and the HN "Who is hiring" thread when it is dominated by robotics posts.

The August 2026 HN "Who is hiring" thread included direct-from-founder robotics posts such as Bucket Robotics in SF. HN's rule that posters must be from the hiring company, one post per company, no recruiting firms, makes it a high-signal channel for founders tracking who is actually staffing up. Watch it monthly.

For GitHub, the repos to watch:

• Physical-Intelligence/openpi
• allenzren/open-pi-zero
• NVIDIA/Isaac-GR00T

Contributors to any of those in the last 12 months are a better shortlist than any keyword search. The catch is that GitHub does not tell you who they work for now, and LinkedIn does not tell you what they committed. Stitching those two graphs together in plain English is where Refolk earns its keep for humanoid robotics recruiting: describe the person ("committed to openpi in 2025, currently in the US, open to relocating") and get the joined view back.

What to change in your JD this week

If you are hiring for physical AI, three edits will roughly double your candidate flow without changing headcount targets.

1. Drop "Robot Learning Engineer" from the title. Use "Research Engineer, Robot Learning" or "Member of Technical Staff, Robotics" instead. You will match how the pool self-describes.
2. Split the JD. One req for sim-to-real (Isaac Sim, MuJoCo, RL), one req for VLA (VLM backbones, flow matching, large demo datasets). Two shorter JDs outperform one kitchen-sink JD every time.
3. Add a paper or repo bullet. "Co-authorship on a VLA or sim-to-real paper in the last 24 months, or contributions to openpi / Isaac-GR00T" screens harder than any years-of-experience line.

The 340% posting growth is not going to slow. Physical Intelligence is still hiring, Skild AI is still hiring, Figure and Agility are still hiring, and DeepMind Robotics has a standing req. The 4,188-person skill pool is not growing at 340%. Whoever sources by skill graph and paper output instead of title will fill roles this quarter. Everyone else will still be running the same broken Boolean in April.

FAQ

Why do so few VLA engineers use the trendy title on LinkedIn?

The field renamed itself faster than the workforce did. Most people training VLA models today were "PhD student" or "Research Engineer" 18 months ago, and updating a LinkedIn title is a low-priority chore for someone deep in a training run. On top of that, frontier labs like Physical Intelligence and DeepMind Robotics use "Research Scientist" or "Member of Technical Staff" internally because that is what recruits PhDs, so their employees inherit those titles. The people who did adopt "Robot Learning Engineer" tend to work at legacy R&D orgs where the label helped with req approval, not at the frontier.

What is the fastest way to build a VLA candidate list from scratch?

Start with author lists from six papers: π0.5, OpenVLA, MolmoAct, CogACT, SmolVLA, and Isaac GR00T N1.x. Dedupe by name and enrich with GitHub handles by cross-referencing commits to Physical-Intelligence/openpi, allenzren/open-pi-zero, and NVIDIA/Isaac-GR00T. That produces a couple hundred names globally. Filter to US or your target geography, then run current-employer lookups.

Should I hire one person for sim-to-real and VLA, or two?

Two. The archetypes rarely overlap in one head. Sim-to-real practitioners come from RL and control (Isaac Sim, MuJoCo, domain randomization). VLA practitioners come from multimodal LLM training (VLM backbones, flow matching, large demo pipelines). If you interview one candidate for both, you will feel like nobody is qualified. Split the JD, run two shortlists, and you will close both roles faster than the merged req would have closed either.

Is $148,000 the right comp anchor for a Physical Intelligence-caliber hire?

No, it is the anchor for the counter-offer, not for your offer. $148,000 is the median US robotics engineer salary as of Q1 2026, up 14% over 2024. That is roughly what Ford, Astrobotic, or a Microsoft applied-research group will pay, and that is who will be trying to hold onto your candidate when you approach them. Frontier VLA offers from Physical Intelligence, Skild, or Figure run substantially higher, especially with equity. Calibrate to beat the counter, not to match the median.