The robotics industry talks a lot about models. Foundation models for perception. Large language models for planning. Diffusion policies for manipulation. The model improvements are real, and they matter.
But when I talk to founders who are actually deploying robots in customer environments, models are rarely the bottleneck. The bottleneck is everything else.
The real bottlenecks
Hardware reliability. Motors fail. Sensors drift. Cables fray. The physical components of a robot degrade in ways that software doesn't. And when hardware fails in a customer environment, you can't push a patch.
Site-specific customization. Every deployment environment is different. Different layouts, different workflows, different edge cases. What works in one warehouse doesn't automatically work in another.
Data collection. Getting high-quality training data from real deployment environments is hard. It requires instrumenting robots, building collection pipelines, and curating data — all while the robot is supposed to be doing useful work.
Customer trust. Customers don't care about model architectures. They care about whether the robot works, reliably, without supervision. Building that trust takes time and requires a level of operational maturity that most robotics startups haven't built.
What this means for builders
If you're building a robotics company, your competitive advantage probably isn't your model. It's your ability to deploy, operate, and improve a physical system in the real world — reliably, at scale, with real customers.
The companies that win will be the ones that out-execute on the messy, unglamorous work of making robots actually work.