Robot Learning Engineer - Stealth Physical AI Startup
The role
You'll lead our core research thesis: that a proprietary stream of real-world, first-person operational data can materially improve how robots learn. You'll define the evaluation framework and reference baseline, decide how our raw data gets structured for machine-learning use, and drive the experimental roadmap alongside a doctoral researcher you'll mentor directly.
This is applied research embedded in a fast-moving company — not an academic lab, and not a pure engineering role. You'll have access to a growing, real-world dataset that doesn't exist anywhere else, and the mandate to test our core assumption rigorously rather than confirm it. Findings get validated internally first; external publication is a possible downstream outcome, not the goal.
What you'll do
- Establish the research framework: pick the task, the success metric, and the comparison baseline — and lock these before testing begins
- Determine how our operational recordings are structured and represented for model training, in a way that also respects our data-privacy architecture
- Lead the experimental program — comparative tests, baseline checks, and unflinching reporting of results either way — in partnership with a researcher you supervise
- Serve as the internal authority on this research direction, including flagging when the underlying assumption doesn't hold
What you get
- Access to a real-world dataset that no outside team has, expanding continuously
- Genuine latitude to disprove the hypothesis — a well-run negative result is treated as a win, not a setback
- Ownership of a research direction you help shape, with a supervised researcher and a dedicated data engineer supporting you
- Direct access to technical leadership and a real say in a long-term strategic bet
Where you'll be in a year
Your evaluation framework and baseline will be locked. Initial comparative results will exist and be reported candidly, whatever they show. The data-structuring pipeline will be specified and operational, and — if results warrant it — a paper will be in progress. Either way, the company will have a sharper, evidence-based view of this data's value than any competitor.
Who you are
- You value getting the right answer over confirming the one you started with
- You can design tests that could genuinely prove you wrong
- You move comfortably between research code and production data systems
- You can mentor a junior researcher without a traditional academic support structure around you
- You can make the case for a research direction to both technical and non-technical stakeholders
Must have
- Practical, hands-on experience with modern robot learning methods — policy learning from demonstrations, generative/diffusion-based control, or vision-language-action approaches — with models you've actually trained
- Background in first-person or video-based learning
- Strong PyTorch skills and experience with large-scale data pipelines
- A track record of research carried through to completion (publication or equivalent)
Nice to have
- Experience with manipulation benchmarks or simulation environments
- Multimodal representation learning
- Exposure to real-world industrial robotics data