Autonomy is scaling fast. As AI performance improves and hardware costs fall, autonomous vehicles are becoming increasingly common in more cities.
Why isn’t local delivery yet autonomous? Because it is fundamentally a more complex challenge.
First, delivery poses a “last ten feet” problem. Unlike ride-hailing, where consumers take themselves to and from vehicles, many consumers want food or groceries delivered to their door, and many merchants don’t want to have to walk around the block to hand off an order. This requires vehicles and software that are nimble enough to handle sidewalks and driveways, but fast enough to get ice cream to you before it melts while navigating street traffic.
Second, local delivery operates on a complex, three-sided marketplace. Autonomy must coordinate busy merchants that need to hand off their products quickly and reliably; consumers with different preferences for how they want their stuff delivered; and the robot connecting the two. Pickups are especially challenging: different merchants want the robot to park in different places. And who loads the delivery?
Third, delivery happens on a dynamic, real-time marketplace. Orders might surge at lunch, but slow at 3pm; demand might peak in certain parts of town, or during holidays like Mother’s Day where we see a spike in demand for flowers; the availability of Dashers constantly changes. This creates a real-time, multi-objective optimization problem across dozens of variables that changes every few seconds.
But if we get these challenges right, autonomy could transform how local businesses operate. Restaurants could expand their delivery radius. Corner stores could compete with major retailers on convenience. Neighborhoods could access a wider variety of local businesses, all while reducing traffic congestion and emissions from delivery trips. It’s infrastructure that could drive real business results for hundreds of thousands of merchants.
DoorDash’s Advantage: Training on commerce at scale
Making commerce autonomous means rethinking autonomy from first principles. We built Dot and our Autonomous Delivery Platform on top of one of the world’s largest delivery networks. We know what supply and demand looks like for our use case: we have billions of data points about effective merchant pickups and consumer drop-offs all in the last 10 feet, and billions of ground-truth commercial outcomes. Was the delivery successful? Was the consumer happy with their order? Did merchants sell more stuff?
We’ve built our technology stack to take advantage of these insights, optimizing autonomy for local businesses and consumer success as well as for driving performance.
Our commerce network is growing rapidly. Consumers can now use DoorDash to order everything from dinner, their groceries, laptops, tennis rackets, cat food, mouthwash, flowers, and much more from local businesses.
As demand rises, and our offerings continue to expand, we are preparing for a future of local delivery powered by Dashers, Dots, drones, sidewalk bots, and other new delivery technologies that help meet merchants’ and consumers’ needs. We expect Dot could grow alongside our network to one day become one of the largest deployments of autonomous technology in the world.
DoorDash Dot: The first commercial autonomous robot to travel bike lanes, roads, and sidewalks — purpose-built for local delivery
Dot is purpose-built for local commerce. It is designed for quick and reliable deliveries, right-sized for the demands of our marketplace, and is engineered for safety at every turn.
We designed Dot to fulfill three criteria:
- A robot design whose size, speed and form factor could be applied to many current DoorDash delivery types
- Driving software that can navigate roads, bike lanes and sidewalks fully autonomously
- A design and tech stack that’s efficient and scalable
Dot is our Goldilocks design that meets these criteria. It is big enough to fit six large pizza boxes, carry up to 30 lbs of cargo and, at 4′6″ tall, be highly visible to other road users — but is small enough to fit through most doors. Traveling up to 20 mph, Dot can ship items faster and much further than sidewalk robots. But weighing only 350 lbs — less than a tenth of the weight of the average new U.S. car — it has a significantly improved safety profile for other road users compared to passenger vehicles.
Dot’s autonomy stack has been architected from inception to take advantage of the latest advancements in AI and autonomous technologies. The perception stack uses a vision-primary approach powered by 8 external cameras providing 360-degree coverage (plus 1 interior camera to ensure delivery quality) supported by 4 inexpensive radar units.
Three high-resolution lidar sensors are currently included for situational awareness, but these are being replaced by inexpensive automotive-grade lidars resulting in a low-cost sensor stack to enable rapid commercial scale.
Dot’s driving software must learn and adapt to the wide range of situations it could encounter during daily operations. Dot is designed to handle busy parking lots, streets, bike lanes, sidewalks, driveways, and paths to enable seamless merchant handoffs and the kinds of deliveries that already take place on DoorDash’s platform every day.
Our stack combines deep learning and search-based algorithms to find a safe, smooth path through a complex world. Deep learning allows us to understand how other road users act, and, in turn, how we should drive in complex situations. The search acts as a safety net and ensures the robot navigates the environment safely, smoothly, and predictably. We train with behavior cloning and reinforcement learning on large, diverse data so the system keeps improving with real-world experience. Our team brings decades of experience shipping safety-critical autonomy systems — and we’re growing fast.
Validation and verification of the autonomy stack is vital to ensure the safety of our robot. We believe we have set the bar for how delivery bots should be tested. We test against a range of edge cases in diverse suburban environments, including construction zones with temporary signage, trucks blocking bike lanes, pedestrians in low light conditions, animals off leash in residential areas, broken traffic lights, school zones, and much more. Our stack is carefully optimized for the communities we operate in.
Scaling a production-grade autonomous technology stack onto a robot of that size is challenging enough. Building a system that can scale to fully autonomous real-world operations — delivering real orders for real consumers while navigating the complexities of real-world driving — compounds that challenge.
Finally, Dot is optimized for operational efficiency. Its front and rear drive modules can be individually swapped and replaced for service or other operational needs, and its fully modular design allows different customized cargos for different delivery use cases: pizza, groceries, and more.
Our Autonomous Delivery Platform: AI to orchestrate autonomous commerce
The technical challenge isn’t just building the vehicle. It’s embedding that new delivery method into the technology platform that orchestrates our marketplace.
Different order types are best served by different delivery methods: not everything is best delivered in a car or on a bike. Dot is designed to scale alongside many other fast-growing delivery methods, including drones, sidewalk robots, and Dashers, to ensure that we can match the right order with the right delivery method at the right time.
That is a complex, multi-objective optimization problem. It requires balancing the characteristics of different orders and delivery methods with consumers’ preferences, traffic conditions, fleet positioning, and many other factors. It means managing the supply and demand of each delivery method available in real time. It means integrating delivery methods from a growing range of suppliers: Dashers, Wing’s drones, Coco Robotics’ sidewalk robots, and of course, Dot. And it must integrate seamlessly into our infrastructure for order processing, our dispatching algorithm, and consumer support, all while maintaining the sub-100ms response times for order evaluation.
We’ve built an Autonomous Delivery Platform (ADP) on top of our existing infrastructure to facilitate this orchestration. Its goal? To dynamically match every order that DoorDash fulfills with the most efficient and effective delivery method.
We’re in the early days of the ADP, rolling it out to cities where we are scaling deliveries by Dot, drone, and sidewalk robot alongside Dashers. We’re building its next generation to support more delivery modes and geographies. Huge technical challenges remain. Calculating estimated arrival times for deliveries made by Dashers alone is a formidable technical challenge at DoorDash’s scale; the complexity of managing multiple delivery methods compounds that complexity. There’s much to build.
But as we add new modes of delivery and they scale across markets, we expect the ADP’s role managing our logistics and marketplace to expand in parallel. To prepare for that scale, we’re experimenting with increasingly ML-powered mechanisms of demand prediction in the ADP, for example. If successful, it can unlock a new frontier for our growth.
While others aim to build end-to-end AI for driving, we’re building end-to-end AI for commerce — where every decision cascades through an AI-enabled marketplace in real time. It’s an exciting time for DoorDash’s growth.
Things we’re building as we scale
As we scale the number of vehicles across geographies, we’re building the next generation of optimization infrastructure. A few technical challenges we’re working on:
We’re continuing to improve our camera-first perception stack to handle more edge cases as we transition from our current sensor suite to cost-effective compact lidars while maintaining or improving the safety of our system.
We’re building a foundation driving model that seamlessly handles the transition between road driving, sidewalk navigation, and precise driveway maneuvering — all within a single inference pass — to master the range of driving contexts Dot faces. And we’re building those models to generalize, handling the infrastructure patterns, local driving behaviors, and environmental conditions that vary from location to location.
We’re scaling RL systems to handle the long tail of delivery-specific scenarios, like navigating around a child’s bike left in a driveway, or safely approaching a front door when a dog is wandering by — scenarios that rarely appear in traditional AV datasets, but are daily realities for delivery.
And we’re building systems that optimize for the assignment and coordination of a multi-modal fleet of many Dots, drones, sidewalk robots and Dashers, all while handling the different dynamics of different delivery options (faster Dot, slower sidewalk robots) and fluctuating demand.
We’re building autonomy for last-mile delivery, across road types, at commercial scale. And we’re building the intelligent infrastructure that makes autonomous commerce possible to support it. Every algorithm, on the robot or in our infrastructure, can impact millions of orders and thousands of small businesses. The technical challenges are hard, the scale is massive, and the real-world impact is immediate. And we’re just getting started.
We are hiring for dozens of open roles in hardware, software, AI, and autonomy engineering, and operations in our DoorDash Labs team to scale Dot and other technologies. Join us.