Robots moving from the laboratory to real industrial applications are far more complex than imagined.

Last year, the field of robotics research indeed achieved many results—progress in VLA, Sim2Real, cross-ontology generalization, and dexterous manipulation were quite solid. But interestingly, the most cutting-edge concerns in academia and industry are completely different. The machine learning teams care about different issues than the companies truly building industrial robots, and there's a huge gap between them that’s hard to bridge.

The main bottlenecks are threefold: First, the data used for training often differs significantly from the real deployment environment; a well-annotated dataset placed on the production line can easily cause failures. Second, research usually focuses on average performance, but industrial applications are most concerned with extreme cases—one mistake can mean significant costs. Third, performance and latency are always in conflict: models that process quickly lack accuracy, while precise solutions take too long to respond. Without resolving these three issues, even the best technical papers are just theoretical exercises.