Starship was founded in 2014 by Ahti Heinla (founding engineer of Skype) and Janus Friis (co-founder of Skype). So far, the investors have backed the startup with more than $100,000,000.
The Starship's robots use the same amount of energy for a delivery that a kettle needs to boil water for just one cup of tea, so they are super efficient.
The robots use sensors and machine learning to understand the environment around them, so they are safe, too. And people like them because of their cute exterior design created by Antti Mäkelä.
Regular cameras provide a 360-degree view and Time-of-Flight (ToF) sensors detect objects and distances in 3D space. The cheap ToF "cameras" are like LIDARs, but instead of lasers they use pulsing LEDs, which is enough on short distances.
The ultrasonic sensors (which we call parking sensors in cars) detect curbs and walls, while a hidden radar detects faster moving objects from a longer distance, like bicycles and cars.
Regular GPS offers 3 meter accuracy, but together with the neural network enabled computer vision and fine-tuned Levenberg-Marquardt algorithm Starship's robots know their location with under an inch precision. Gyros round up the fine sensors in Starship robots.
The controllers of the cameras, radars and motors are made by Starship. Starship says the cost of their robot is about as much as a high end laptop. Most of the costs account to R&D. The energy storage is the most expensive hardware component, but generally the components are cheap (computer around 100€, cameras around 5€ a piece).
The 1.26 kWh battery is the heaviest component, pushing the robot's weight to 35 kg/75 lbs empty. The heavy weight is not all bad - it might passively serve as an anti-theft measure the moment someone lifts the robot!
The electrically-powered robots charge themselves with the in-house designed chargers over the night in 6 hours and then they are ready to travel around 20 miles/30 km the next day. Their travelling speed is electronically limited at 4 mph/6 km/h in order not to pose any threat to the pedestrians and to make the robots behave more human-like. That increases social acceptance.
To make the experience even nicer to humans, the robots speak through the speaker. The latter can also be used by the operator. Naturally, there's also a microphone. The robot even has sirens that activate when it is picked up or tampered with. That hasn't happened, though - after all, in addition to its heavy weight, the robot has GPS, cameras and live internet connection contributing to its security.
In 2021, 150 Starship robots a month were handmade in Tallinn, Estonia.
Users use a smartphone app to order their food from a restaurant or groceries from a store. While waiting for the delivery to arrive, they can monitor robots' journey on their smartphones.
There are 2 million lines of code in the robot plus 2 million lines of code in the servers meaning there's as much action in the cloud as there is inside the robot.
There are around twenty five C++ developers working on the robot's embedded software and roughly the same amount of devs working on the software in the server.
Node.js was initially used in the backend, but they are switching to Go and maybe even Elixir (there's one service already written in it). In addition to these programming languages, Starship's software stack includes C, Python, Rust, React, React Native, Redux, PostgreSQL, MongoDB, Linux, Terraform, Ansible, Jenkins, AWS EKS Kubernetes, AWS EC2 Elastic Compute Cloud, AWS SQS message queuing, Kafka, GraphQL, Prometheus and Grafana, Loki, Linkerd, Jaeger, Apache Spark, Databricks (irreplaceable!), Tableau, Redash, Airflow and more.
In autonomous driving, latency is the enemy. To overcome the slowness of 4G, split-second decision software is naturally in the robot. For other decisions - where more time can be spent - cloud computing is used.
Starship has around 150 vital microservices in the cloud. For example, the 3D map server, order service, route calculation, crossing the street feature, opening the lid feature ... you name it.
The embedded software is designed to detect what the traffic lights show as autonomous driving would not be possible without it.
In cases when the robot encounters an unexpected situation, it stops and an operator is automatically informed. Operator checks what the cameras show, and usually allows the robot to carry on. In very rare cases will the operator take over to manually drive the robot, for example, over a brand new crossroad that the Starship robots have not seen yet. Every robot collects information and it is shared with the other robots over the cloud within 24 hours.
The robot is recording video with its cameras which means that every robot generates terabytes of data every day. Out of this vast amount only parts of the video are sent to the cloud - in the case of a mapping or safety issue - and the rest is deleted in 48 hours.
Because of the complexity in Starship software, Starship hires almost exclusively senior developers. Based on the toughts of the founder and CTO of Starship, Ahti Heinla, the formula of a good developer seems to be something like this: a good developer is smart, can prioritize, can predict errors, delivers fast, has leadership skills and is experienced.
Starship Technologies' business development HQ is in USA, app design in Finland and the rest in Estonia including R&D, manufacturing, operations, data team, AI and software development.
Starship's delivery service was launched in 2018 in UK and in 2019 in USA. The service is being tested in other countries, too. The number of robots in operation in 2021 was over 1000. Based on usage data, the average distance of one delivery round trip is 1.4 miles/2.2 km.
Starship's robots service other types of businesses, too - for example, a chemical plant.
The first 1,000,000 deliveries were completed in January 2021.
To protect the intellectual property, Starship does not publish anything about their research and development. There are companies which have copied the Starship robot, but the highest value is in the software, especially in the self-learning AI part.
What seals the success is that people call Starship's robots "cute". That is a quality of a living being. And being like a living being, might be the ultimate goal for robotmakers.