Why I Think Data Infrastructure Is the Real Robotics Bottleneck

The more time I spend thinking about robotics and multimodal AI, the more I suspect that one of the biggest long-term bottlenecks is not just models. It is data infrastructure.

A lot of progress in AI has been accelerated by large, usable datasets and the ecosystems built around them. In robotics, that problem is harder.

The data is more expensive to collect. The environment matters more. The signals are more heterogeneous. And the information that matters is often not fully captured in standard formats.

That is a big part of why I keep returning to dataset design as a serious technical problem rather than a supporting task.

Work like SCOUT++ pushed me further in that direction because it made the value of better structure very obvious. If grounded language, perception, and action data are not aligned in a usable way, a lot of downstream work becomes harder than it should be.

If we want more capable embodied systems, I think we will need better ways to collect, structure, align, and reuse grounded interaction data at scale.

That belief is shaping a lot of how I think about future work, including the possibility of building around robotics data more directly.

Because in the long run, the teams that control high-quality grounded data may end up shaping a large part of what becomes possible.

---

Originally posted on LinkedIn.