Robotics labs hype autonomous delivery every day, but I have one question: who’s going to tackle the rough roads in the countryside? Downtown grids, with clearly marked lanes, orderly traffic lights, and wide-open intersections, make it easy to train a system and produce a slick demo video. But once you get to suburban dirt roads, nameless lanes, and house numbers all over the place, AI is completely lost. Try getting it to pass another vehicle on a two-lane rural road: there’s an irrigation ditch on one side, bamboo mats spread out to dry grain on the other, and a three-wheeler pulled over up ahead. The road markings have long since worn away, and the map doesn’t have a single decent POI. What can it base its decisions on? This isn’t a technology problem. It’s a data problem caused by cutting corners.

A lot of autonomous delivery “scorecards” look the same: university campuses, gated communities, urban pilot zones—fixed routes, clean environments, clear right-of-way. A few thousand deliveries a day sounds impressive. But public datasets and data-collection vehicles mostly focus on structured urban roads. In datasets like Cityscapes and nuScenes, lane markings, lanes, and traffic signs are the main features. In the countryside, though, there are narrow bridges, gravel roads, muddy potholes after rain, piles of earth dumped temporarily, guy wires on utility poles, and chickens, ducks, and dogs darting into the road—all long-tail scenarios. As of the end of 2023, China’s rural roads stretched more than 4.6 million kilometers, yet rural roads make up a tiny fraction of the training data for many delivery robots. Missing house numbers make things even worse: in a city, “No. 88, XX Road” can get you to the door. In the countryside, directions are more like “the third house past the little shop at the village entrance.” Can AI make sense of that?

The best low-cost opportunities are rarely in the spotlight. The real challenge for autonomous delivery isn’t dropping off coffee outside an office tower; it’s handling the last mile for parcel deliveries in towns and villages. The real demand is hidden in the places everyone overlooks: few delivery hubs, scattered packages, long trips for older people to pick up parcels, and van runs that aren’t worth the cost. But these settings have poor road conditions, little data, and expensive labeling—and even the results might not look impressive. So companies would rather keep looping around in pilot zones. Growth without volume is a sham; sooner or later, you have to make up the volume. No matter how fast the mileage, order counts, and number of communities served on a launch-stage slide keep climbing, without long-tail data from rough rural roads to back them up, the louder the hype, the less real-world traction there is. Maybe I’m wrong. I’ll wait to be proven wrong.