Founding ML Engineer

What you'll be doing

• Implement and train world model, VLA and reinforcement learning architectures on our in-house and customer datasets.
• Read and discuss papers, decide which methods and ideas are worth implementing and should define our research agenda.
• Run models on factory floors next to the people using them, reason about failure modes, and exercise judgment about whether a problem is best solved with more data, better models, or a user-interface change.

What we look for

• A demonstrated willingness to go further than most people would consider reasonable to reach your goals.
• A strong grasp of probability, statistics, and ML fundamentals. You should be able to explain the intuition of KL divergence on a napkin, and be familiar with concepts such as flow matching or PPO.
• You're comfortable writing code in Python and PyTorch, but also know how to orchestrate agents so you rarely have to.
• Strong grounding in CS, ML, physics, maths, or robotics, e.g. master's, very strong bachelor's, or equivalent experience. We care more about how fast you learn than what you already know. Most of what we do is still being invented.
• You can get an idea out of your head and into someone else's intact, starting from what they know, not what you do. We'll be learning from each other constantly, and a good idea that doesn't make it across gets rebuilt from scratch or dropped entirely.
• You argue hard for your position and drop it the moment someone shows you something better.

Plus points

• Prior experience training and/or finetuning/post-training LLMs, VLMs, VLAs, or World Models.
• Experience with reinforcement learning. Reliability is where the value is, and RL is essential to getting there. Classic methods like PPO don't transfer directly to the models we work with, but fluency with the concepts helps a lot.
• Published papers in ML. We care less about the paper's topic than that you've turned a raw idea into structured experiments and drawn defensible conclusions.
• Experience with autoresearch.
• Experience working with CAD and/or robotic hardware.
• A demonstrated interest in sharing what you learn via personal blogs, videos, social media, or a form of your choice. We're not an open-source company, but we build in the open. Sharing what we learn and helping the robot-learning ecosystem is a win for everyone.
• Fluency with at least one major European language (DE, FR, IT, ES, PL). We will be spending time in factories where not everybody speaks English.

What we don't offer

• Remote work. We're in person in Zurich, 6 days a week and work on most public holidays. If you're seeing the same developments in robot learning as we are, it will be obvious why.
• A clearly defined role. Fixed roles are for solved problems, and nothing we work on is solved yet.
• Glamorous work. Work will include days, maybe weeks, on factory floors in towns you didn't know existed.
• Overnight success. Success isn't a moment, but a lot of unremarkable Tuesday nights, and our trajectory will be as smooth as a step function. We're not optimising for views and this isn't a place for people who need quarterly validation. We're signing up for the Tuesdays.
• Many meetings or detailed processes. We're a jazz band, not an orchestra.