Robot Learning Engineer
At a glance
Mid-level AI Research role at Rohlik. Prague · full-time.
Pay not stated AI Research salary
aimlroles summary, based on the employer's posting.
What you'll do
- Set camera, mounting, calibration and episode-quality rules, then check them against conditions on the warehouse floor
- Define how hand-pose accuracy is measured and address failures caused by gloves, occlusion and cold warehouse conditions
- Build the evaluation process that filters recorded hours for policy training and keeps the quality threshold steady as the dataset grows
- Train open robot-learning models on collected recordings and evaluate their first tasks
- Review published methods, reproduce key results and use what you learn to guide capture decisions before spending
What you bring
- Strong PyTorch and computer-vision skills, including practical experience with 3D geometry
- Working knowledge of camera calibration, pose estimation and SLAM, including debugging extrinsic calibration
- Ability to reproduce research results and familiarity with robot-learning policies and the data they need
- Experience training models on data you collected yourself and spotting issues that emerge during training
Who this fits
This mid-level role suits an engineer who wants to own dataset quality and early robot-policy training rather than work only on other teams’ models. You’ll be the sole machine-learning voice on a small team, working alongside an operations lead and a data engineer. The role is hybrid in Prague.
From the employer
Why this role is exciting
Rohlik is the leading Central European e-grocer. More than a million customers buy their groceries from us through Rohlik.cz, Knuspr.de, Kifli.hu, Gurkerl.at and Sezamo.ro, choosing from over 17,000 items, delivered within a couple of hours in 15-minute windows.
We are building the autonomous grocer. Maia, our AI assistant, already talks to customers and builds their orders; agents are moving into our buying and logistics; vision models are learning to watch quality. The next frontier is physical: robots taking over more and more of the work in our fulfilment centres, until fresh food moves from producers to households at a cost nobody else can match.
We are setting out to record how our own people handle groceries and to turn those recordings into a training corpus for manipulation policies. Our warehouses already log what happened on every pick and every quality check, so the recordings can carry labels most datasets never get. Making that hold is part of the job.
You decide what makes an hour of footage trainable — before we bank hundreds of them.
What you will own and deliver
The capture spec. Cameras, mounts, calibration, and what disqualifies an episode. You write it before we buy the fleet and enforce it after. When the spec and the floor disagree, you go to the floor and find out which one is wrong.
Ground truth. Hand pose is the signal that transfers from human video to robot policies, and warehouse reality — work gloves, occlusion, cold halls — is exactly where the published models struggle. You own how we measure that and how we close the gap.
The evaluation harness. The corpus only counts if a policy can train on it. You build the harness that decides which hours make the bank, and you hold the trainability bar as the volume grows.
The first training runs. Once hours bank, you post-train open robot-learning models on our data and run the first task evaluations.
Judgment before spend. The field moves monthly. You read what is published, reproduce the claims we depend on, and steer our capture before the money is spent, not after.
The scope
This is a seat on a small, newly created team: an operations lead who runs capture on the floor, a data engineer who owns the pipeline, and you: the machine-learning voice in the room. This is not a research-scientist role, and no publication record is required. The job is to make a corpus trainable and to train the first policies on it.
What we are looking for
PyTorch, and computer vision with real 3D geometry. Camera calibration, pose estimation, SLAM fundamentals. You have debugged an extrinsic calibration at least once in your life and know why it drifted.
You can read a paper and reproduce it. Our pipeline is assembled from published work, and your job is knowing which claims survive contact with a chilled warehouse hall.
You know what policies train on. Familiar with the current robot-learning stack and policy classes well enough to know what the training side will demand of the data before the data exists.
Dataset instincts. You have trained on data you collected yourself, and you know the failure modes that only show up at training tim..
Comfortable as the only ML voice. You make the call, write it down, and revisit it when the pilot data says otherwise.
AI is how work gets done here. Our engineers use Claude Code and Devin daily, and we expect you to direct agents for the unglamorous parts — harness code, data plumbing, reproduction scripts — while keeping the quality bar exactly where it was.
Particularly relevant. Hand-pose estimation or egocentric video in production; multi-camera rigs and time sync; experience post-training robot-learning models; ROS 2.
What success looks like in 90 days
You have written the capture spec and we set up the first rigs against it. Your pilot gates have numbers, the first hours have passed your harness, and you can tell the people who sign the budget, with data, what the next hundred hours should look like.
Are we a good fit?
Take this role if you would rather own the spec that decides whether a dataset becomes an asset or a write-off than tune someone else's model on someone else's data, if a fulfilment centre at 6 a.m. sounds like a lab to you, and if you want the shortest possible line between your technical judgment and a capital decision. Read our culture code and are we a good fit?.
What we offer
The physical-AI side of the autonomous grocer, with the capture spec and the first training runs yours from day one
A data asset few teams in the world get to build, in fresh groceries, a domain as hard as manipulation gets
Off-the-shelf hardware, a managed cloud estate, and a mandate that has already started
Benefits — see what we offer
Are you in?
Instead of a cover letter: pick a published human-video-to-robot-policy pipeline you rate, and tell us where it breaks on a warehouse floor where every hand wears a work glove.