If you've ordered takeout from Uber Eats in Austin or Miami, there's a chance your food was delivered by a robot. What looks like a last-mile delivery is also a training exercise for the next generation of autonomous systems.
Companies including Coco Robotics and Avride are using commercial deployments to generate the operational data needed to improve autonomous systems and expand delivery into more cities.
What cities teach robots
For Zach Rash, co-founder and CEO of Coco Robotics, every commercial deployment contributes to the dataset used to improve future AI models.
"We have millions of hours of driving in these sorts of environments," Rash said. "That can be used to train a really reliable self-driving model on the robots."
City streets are simply too unpredictable to reduce to a fixed set of rules.
Coco’s delivery robot on its way. Credits: Coco Robotics
"The reality is you're just always surrounded by people on e-bikes flying by you and cars and people walking and people coming in and out of stores," he said. "Creating those rules is impossible."
When robots encounter unfamiliar situations or become stuck, operators can step in to help: that human intervention becomes data too.
"The best way to create that dataset is to provide actual value, economic value, so you can continue to collect that data and iterate from there," Rash said.
The biggest difference from self-driving cars
Avride has reached a similar conclusion from a different starting point. Rather than building delivery robots first, the company adapted technology originally developed for autonomous vehicles.
"The overall goal for us internally is to reuse as much as possible from the bigger brother, from the car project," said Roman Nefedov, head of delivery robots at Avride.
Much of that autonomy stack transferred intact. The robots share localization, perception and some prediction technology with Avride's self-driving cars, allowing them to identify objects and anticipate how the world around them is likely to evolve.
The biggest difference lies in planning. According to Avride, cars operate in a structured environment of roads and lanes, where other road users are expected to follow a set of rules. Sidewalks, by contrast, are a completely unstructured environment of pedestrians, cyclists, dogs, street furniture, and temporary obstacles.
Delivery robots on a mission. Credits: Avride
"The algorithms for feeling the world are the same," Nefedov said. "The algorithms for planning what to do next, they're completely different."
Understanding the world also means knowing precisely where the robot is within it. GPS can identify the correct street, but often lacks the precision needed to locate the exact front door, apartment entrance, or collection point.
To address that challenge, Coco partnered with Niantic Spatial, whose Visual Positioning System combines geospatial AI and computer vision to localize robots when GPS alone lacks sufficient precision.
"The maps that you need for robots are quite different from those for a person," Rash said. "The good thing about robots is you can tell us exactly where to go, and we will go exactly there. The bad news is you might not have picked the right spot."
Engineering for the real world
Operating hundreds of robots every day requires different engineering. One challenge is deciding what intelligence belongs on the robot and what can be handled elsewhere.
Human oversight remains an important part of that process. Yulia Shveyko, head of public affairs at Avride, said the company's onboard perception and prediction systems can recognize individual objects, but have a more limited understanding of the overall scene.
To add that context, Avride uses a cloud-based vision-language model (VLM) that receives camera snapshots from robots every few seconds and flags situations that may need human monitoring. The VLM does not control the robot, avoiding the latency and connectivity risks of cloud-based control.
"The VLM itself does not control the robot; it does not send any commands," Shveyko said. "It works as an alert system."
That can help determine whether a first responder is actively dealing with an emergency, for example, or identify wet cement at roadworks rather than simply recognizing cones and barriers. When the system flags a potentially difficult situation, a remote assistant can monitor it and, if necessary, send instructions for the robot to execute locally.
As fleets grow, managing the data they generate becomes an engineering challenge. Every robot continuously records video, lidar, telemetry and other sensor data, producing up to a terabyte of information during a day's operation.
"The engineering team always wants more logs because data is just the fuel for improvement," Nefedov said.
Uploading all that data is expensive. Rather than transferring everything over mobile networks, Avride selects the most valuable moments for future model development. Those decisions matter more as commercial fleets expand.
Not SOS, just ZZZ
Energy efficiency creates another set of trade-offs. Autonomous delivery robots consume roughly the same amount of power as a walking human, but around half of that energy goes to sensors and onboard computing. Every watt spent processing AI models is a watt unavailable for extending operating range.
Coco delivery robot in the snow. Credits: Coco Robotics
For Nefedov, those engineering compromises distinguish commercial deployments from technology demonstrations.
"Energy is a huge concern for us, and that's what differentiates a demo from real work."
That focus extends to surprisingly small details. During quieter periods between lunch and dinner, robots shut down as many systems as possible to conserve battery power. Even the LED display was redesigned to minimize energy consumption after pedestrians mistook stationary robots for broken machines. Instead of switching the display off completely, the robots simply show "ZZZ" to indicate that they are sleeping while waiting for the next delivery. Shveyko said reports from people concerned that stationary robots were stuck have since dropped dramatically.
Learning in the real world
For Coco, expanding beyond Los Angeles required more than retraining AI models.
"When we met with the Finnish regulators, their comment to us was that these winters took down the Russian army," Rash recalled. "Are you sure you want to launch here?"
Operating in those conditions required heated cameras, weather-resistant hardware, and new operational data for future deployments.
Avride found that seemingly simpler environments could prove equally challenging. Nefedov said university campuses, despite having fewer vehicles, often proved just as challenging because dense pedestrian traffic demanded a different style of planning.
Coco and Avride have adopted different technical approaches but describe the same underlying dynamic: commercial deployments generate operational data, operational data improves AI models, and better models enable broader commercial deployments.
Finnish winters, crowded campuses, a dog that wanders into frame: none of it can be simulated away. The city still refuses to behave, but it's the only real teacher either company has.
Top photo: Delivery robot on the streets of Austin, USA. Credits: Avride
Avride is part of the Nebius group, which owns The Infinite Loop - Ed.
