Research Engineer, Robotics Foundation Models

Qualia is the company that brings foundation models into the real world: data collection, fine-tuning, inference, and on-device deployment, end to end. We're a small, technical team in Copenhagen and Stockholm building this for manufacturers, in partnership with top companies. Qualia has raised roughly USD 10M to pursue the hardest problems in robotics.

We're hiring a research engineer to push the model side of our research. VLAs, WAMs, that work on real customer data. RL that closes the gap on precision tasks. World models, UMI, EgoData, that let us iterate faster than collecting another 1,000 episodes of teleop.

This is not a research role at a lab. This is a research engineer role at a startup. Your work ships to customers, runs on robots, and gets judged by whether the task succeeds.

What you'll work on

VLA fine-tuning at the edge of what's possible

VLAs on customer and internal data. Make them work where the base model fails.

Build the recipe. Subtask decomposition, stage classifiers, relative actions, action chunking, normalization. Everything that turns "we got 30% success" into "we got 99% success."

Push the data pipeline upstream so that fine-tuning is repeatable across robots and tasks, not a one-off.

RL on top of VLAs

Residual policies for high-precision sub-movements (key insertion, peg-in-hole, force-sensitive contact).

RL-driven training loops on top of VLAs and WAMs.

Sim-to-real where it accelerates a real customer task, not for its own sake.

World models and synthetic data

Build the workflow that turns a small amount of real teleop into a much larger training corpus.

Decide what's worth using in our platform and what isn't.

You'll own model side end to end: data, training, inference, evaluation, iteration.

Who we're looking for:

Required

Strong research track record OR strong applied robotics ML track record. Either path works.

Hands-on with at least one VLA stack. Trained one, evaluated one, debugged one. Not just read the paper.

Comfortable with the main ML frameworks. Comfortable on Linux. Comfortable on a real robot (or eager to be).

Comfortable with messy data and the realization that 80% of the work is making the data not lie to you.

Comfortable shipping. Models in production beat models in arxiv at this company.

Bonus

Real teleoperation or data collection experience

RL beyond toy environments

Experience with sim-to-real, world models, or video-conditioned generation

Familiarity with VLAs

You've shipped a robot policy that a non-engineer used and didn't break

What we don't care about

Whether you have a PhD or you don't

Whether you've worked at a famous lab or you haven't

Whether your background is industry, academia, or self-taught

We care that you've done the thing.

How we work

Small team, fast loop. Around 10 people. Repos move fast. No PR purgatory.

Real customers from day one. You'll talk to pilots directly.

Strong opinions, loosely held. We argue, we ship, we update. You'll be expected to push back when we're wrong.

Hardware is real. Open Arm, R1 Light, Flexiv's, Thors. Not just sim.

What we offer

Competitive salary + meaningful equity (early-stage)

Working on robotics foundation models with a team that ships

Direct line to founders and customers

Copenhagen office, Stockholm satellite opening end of summer

Real budget for compute, hardware, and other things that matter

The usual good stuff

Apply here and or e mail fabian@qualiastudios.dev with whatever best represents your work and/or a resume.