The next generation of autonomous robots will be defined by data. Just as LLMs were trained on internet-scale text, embodied AI will require enormous amounts of real-world demonstrations to learn how humans interact with tools, parts, and complex manufacturing processes. At Syn2core, we’re building exactly that foundation. By capturing multimodal, egocentric data from factory workers at scale, we’re creating one of the richest datasets for industrial robot learning — unlocking the future of autonomous manufacturing.
To make this possible, we’ve built an AI copilot for smart glasses that guides workers through assembly and inspection in real time. Already deployed with leading automotive manufacturers, the system improves quality, productivity, and traceability while continuously generating the high-quality training data that tomorrow’s robots will learn from. If you’re excited about embodied AI, robot learning, multimodal perception, and turning real factory data into autonomous capabilities, you’ll help build the bridge between today’s workforce and tomorrow’s intelligent robots.
The role
You will own our robotics research thesis: that egocentric factory data measurably improves robot learning. You define the benchmark and baseline, design how our data is encoded for robot-learning use, and run the experimental program with a PhD student you supervise day-to-day. This is research inside a delivery-driven startup: real, growing data no lab has, and the rigor to turn it into a defensible result.
You’ll work directly with the CTO and alongside the Data Engineer who builds the research pipeline to your spec. Internal validation first; publication only if the result supports it. A clean negative result is a deliverable here, not a failure — we diagnose hypotheses, not people.
What you’ll do
- Frame the science — define the benchmark and baseline: which robot-learning task, which metric, and what “improvement” means, frozen before experiments start
- Design the data — how egocentric factory recordings are encoded and represented for robot-learning consumption, jointly serving our privacy-preserving storage format
- Run the program — ablations, baselines, honest negatives included, together with our PhD student
- Own the thesis — be the company's expert voice on it, including telling the CTO if it's wrong
What you get
- A dataset no lab has: real egocentric manipulation data from factory floors, growing daily
- The freedom to falsify — negative results are respected deliverables, not career risks
- A research agenda you define, with a PhD student and a data engineer building alongside you
- Direct line to the CTO and real influence on where the company's long-term bet goes
Where you'll be in 12 months
The benchmark and baseline are frozen. The first ablation results exist and are honestly reported — positive or negative. The encoding pipeline is specced and running, and if the results support it, a publication draft is underway. Whichever way the evidence points, the company knows more than any competitor about what this data is worth.
Who you are
- Scientific honesty over thesis-confirmation — you'd rather kill the hypothesis than fool the company
- You design experiments that can actually falsify something
- You bridge research code and production data pipelines without friction
- You can supervise a PhD student without a professor's infrastructure behind you
- You can explain research value to a CEO and an investor, not just a reviewer
Your experience
Must have:
- Hands-on robot learning: imitation learning, diffusion policies, or VLA models — trained, not just read about
- Egocentric or video-based learning experience
- Strong PyTorch and large-scale data handling
- A publication record, or equivalent evidence of research carried to completion
Ways to stand out:
- Manipulation benchmarks and simulation suites
#J-18808-Ljbffr
