RetailNews
DoorDash Tests Its Delivery Robot on Millions of Simulated Miles. A Toy Dog Handles the Rest.
"For us, the challenge is scaling."

The brief
- DoorDash is utilizing cloud simulation via Amazon Web Services to train its delivery robot, Dot, for real-world scenarios.
- The robot logs sensor data and decision-making processes, which are analyzed to improve its performance in various environments.
- Validation of Dot's behavior occurs primarily in simulation before confirming results with real-world driving tests.
DoorDash is using cloud simulation to teach its delivery robot how to handle the road. DoorDash Labs, the company's robotics and automation R&D division, uses Amazon Web Services to train and test key parts of the AI behind Dot, a small red delivery robot, according to an AWS feature published on November 12, 2026.
The brief
- DoorDash is utilizing cloud simulation via Amazon Web Services to train its delivery robot, Dot, for real-world scenarios.
- The robot logs sensor data and decision-making processes, which are analyzed to improve its performance in various environments.
- Validation of Dot's behavior occurs primarily in simulation before confirming results with real-world driving tests.
"For us, the challenge is scaling," said Janek Hudecek, senior principal autonomy engineer at DoorDash. "At scale, a rare event stops being rare."
Dot logs what its sensors see, every decision its software makes and every time a safety operator steps in. At the end of each day, the data is uploaded to Amazon S3 and indexed with Amazon Redshift, where engineers search for moments when Dot did not behave as expected.
The perception team uses Amazon SageMaker to train updated vision models when Dot encounters new environments.
Every change is then tested against the equivalent of millions of miles of simulated driving on Amazon EC2, across hundreds of GPUs in parallel, before it reaches a real robot.
"Nearly all our validation happens first in simulation, and then we use real driving to confirm that what the robot learned in simulation holds true on the street," said DoorDash product manager Jordan Mesch.
Passing standard tests was not enough. Dot passed an industry assessment for stopping when a pedestrian steps out between parked cars, but the team later found it was braking too cautiously whenever its view was blocked.
"The most hesitant robot on the road isn't the safest one," Hudecek said. The team rebuilt the feature around a probabilistic model, so Dot now weighs how likely it is that someone steps out and adjusts its speed within set limits.
Final checks happen on a test track, where engineers douse Dot with water, build rumble strips and pull a stuffed dog on a skateboard into its path.
Dot runs on DoorDash's Autonomous Delivery Platform, built on AWS, which matches each order with the best delivery option.
DoorDash Air, its drone system, will begin pilot deliveries in Northern California with partners including Chipotle and Popeyes. SmartScale, its in-store order-weighing system, has reduced missing-item claims by up to 30%.
Human Dashers remain at the center of the network. "There's no one solution for delivery," Mesch said. "In markets where demand outstrips the number of Dashers, it might help to have robots supporting them."
Join the discussion
0 answers
People who run these systems answer here. Your verified role is shown, never your employer unless you choose.
Sign in with LinkedIn to answer
Read next
Chaos Predictions — every Monday
Where enterprise AI is heading next, read from the week’s deployments across banking, pharma, manufacturing and retail.
Free. One email every Monday. Unsubscribe in one click.












