Building an Autonomous Delivery Experience with DoorDash Co-Founders Andy Fang and Stanley Tang
In a Nutshell
DoorDash is building agentic commerce through natural language ordering and an autonomous delivery platform that combines human Dashers, DOT robots, drones, and AI across all modalities. Their 10 billion deliveries provide unmatched real-world data for training and operations, with Phoenix already running fully autonomous L4 deliveries for two years. The core strategy prioritizes customer use cases over technology-first approaches, enabling a multimodal fleet that will scale demand rather than reduce human workers as delivery becomes more efficient and affordable.
These notes were generated by AI and may contain inaccuracies.
Andy Fang and Stanley Tang discuss how DoorDash enables natural language requests for food and groceries. The conversational experience emerged after an earlier focus on voice interfaces failed to gain traction. The natural language approach allows users to express nuanced restaurant discovery queries and grocery tasks without keyword optimization or extensive research.
On the restaurant side, 50% of trajectories using Ask DoorDash involve ordering from places customers have never ordered from before. On the grocery side, basket sizes are 40% larger, with users taking photos of their fridges, conducting meal planning with dietary constraints, or reordering usual items more easily than through traditional interfaces.
DoorDash incorporated external world knowledge into the experience, including trending restaurants from internet sources, forums, and social media. This addresses knowledge cutoff limitations in models and helps customers discover what's currently popular.
The founders speculate that a DoorDash created today would likely be more agentic-first. They note there is now more agent traffic on the web than human traffic, suggesting future DoorDash experiences will need to accommodate this shift.
One example involves cameras monitoring office pantry shelves that can automatically trigger DoorDash queries when items run low. Early experimentation with their CLI aims to reduce friction for agents to participate in the ordering experience.
A described use case involves hosting extended family dinner every Sunday, polling attendees for dietary restrictions and preferences, then placing orders. Similar complexity exists for office lunch ordering where managers must accommodate various allergies and preferences while ensuring timely delivery.
DoorDash has been investigating robotics and autonomy since 2018, before it was clear these technologies would become significant. The company began as a skunk works project with Andy and half an engineer's time, initially focused on exploration rather than building hardware.
DoorDash's approach mirrors their founding story with politely.com, which started with eight PDF menus and a Google voice phone number. All major initiatives, including autonomy and AI, begin as experiments focused on customer problems before scaling into full programs.
For several years, DoorDash partnered with existing robotics companies rather than building their own technology. They worked with sidewalk robot companies and robo-taxi players. Three key learnings emerged:
- Validated that autonomy was inevitable, not a question of if but when
- Identified the extensive infrastructure, ecosystem, and operational requirements beyond just deploying robots
- Revealed the importance of building toward specific use cases rather than retrofitting technology
DoorDash built an autonomous delivery platform encompassing products, technology, APIs, and dispatch systems needed for a post-autonomy world. This includes merchant integration methods and consumer experience design.
The founders observed that many autonomy startups built technology first then sought problems to solve, often resulting in poor fit. DoorDash's approach prioritizes the customer use case first, then works backwards.
Sidewalk robots traveling 2-3 mph proved too slow for DoorDash's average 3-5 mile deliveries with 15-minute target times. Robo-taxis designed for passengers presented different issues: 4,000-pound vehicles unnecessary for carrying burritos, and pickup/dropoff requirements that differ significantly when transporting goods versus people.
Robo-taxis can drop passengers half a block away since people can walk, but packages cannot. Food must be picked up directly at merchants and delivered to specific locations like driveways or porches.
DoorDash developed DOT, an in-house autonomous delivery robot weighing 300 pounds, traveling up to 20-25 mph, designed for bike lanes and roads in addition to sidewalks. The form factor sits between sidewalk robots and robo-taxis, appropriate for dense suburban deliveries where most DoorDash activity occurs.
DoorDash has conducted 10 billion deliveries, providing extensive operational insights. The company handles over 3 billion deliveries annually, with each delivery unique in shape, size, geography, and requirements. This diversity spans downtown San Francisco versus Dallas, snow conditions in Helsinki, pizza versus ice cream, and various merchant types from McDonald's to mom-and-pop shops.
DoorDash launched Tasks, where Dashers help collect data points to train world models. The founders note significant investment required for physical world applications compared to software solutions like ChatGPT, with pressure to deliver better value than current alternatives in quality and cost.
DOT operates in Phoenix and Tempe, conducting fully autonomous L4 deliveries for nearly two years. The experience resembles Waymo in San Francisco.
Robotics success requires handling any object or environment, not just cherry-picked demonstrations. Real-world data provides the most realistic training data, giving DoorDash advantages through their delivery volume and variety.
DoorDash is pursuing a multimodal delivery approach rather than attempting to solve all delivery scenarios with a single method. The strategy involves deploying different modalities based on specific use cases and conditions. DoorDash Dot handles 3 to 5 mile suburban deliveries from strip malls, with Phoenix serving as the initial market due to its dense suburbs where destinations are far enough apart to make this approach viable.
Different scenarios call for different solutions: rural areas with poor road infrastructure and lightweight orders may use drone delivery, while complicated multi-step grocery orders requiring stairs and pick-and-pack operations continue using human Dashers. The approach allows phasing in modalities over time rather than requiring an all-or-nothing implementation.
From both consumer and merchant perspectives, the experience remains identical regardless of delivery method. Customers continue using the same app to access all services, while merchants maintain a single integration point. Once integrated with DoorDash, merchants gain access to Dashers, drones, autonomy, and AI tools and products.
This ecosystem approach for local commerce creates significant competitive barriers, as replicating DoorDash's infrastructure across 40-50 countries with diverse merchant types represents a substantial challenge.
DoorDash faces competition for robotics and applied AI researchers who often gravitate toward general AI problems rather than specific use cases. The recruitment pitch emphasizes real-world deployment over laboratory work: candidates can either work on prototypes and demos in academic settings or ship actual products that impact the real world.
DoorDash Labs and AI efforts focus on pragmatic, practical approaches rather than moonshot ideas. The emphasis is on shipping products that interact with the physical world, enabling iterative learning from real-world deployment rather than theoretical research.
DoorDash has conducted autonomous deliveries in Phoenix for over two years, achieving fully autonomous L4 operations last year. The transition from demonstration to scaled fleet operations revealed numerous edge cases not apparent in controlled environments.
Real-world scaling challenges include environmental factors like dirt covering camera sensors, leaves affecting wheel traction and torque requirements, and extreme braking situations where regenerative braking systems can overpower batteries. Operational considerations encompass depot setup, maintenance, battery recharging, and manufacturing supply chain management.
Boot-up processes that worked for small demonstrations became significant bottlenecks at scale, with Jenkins scripts taking 30-45 minutes per robot and failing frequently when managing hundreds of robots daily.
GPS pin accuracy becomes critical for autonomous vehicles, as robots cannot improvise like human Dashers when locating specific storefronts, front doors, or gates within apartment complexes. DoorDash possesses unique data from historical human Dasher drop-off locations that doesn't exist in Google Maps or other mapping services.
This data advantage represents genuine value for robotics applications, as it provides real-world distribution information that cannot be simulated or imagined by engineers.
Three primary scaling components exist beyond the initial autonomy breakthrough: operational scaling across different markets with varying restaurant behaviors and integrations, fleet management and interface layer development, and hardware considerations including supply chain, component reliability, and manufacturing at scale.
DoorDash partnered with Also, a micromobility company spun out of Rivian with RJ Scaringe as board founder and chairman, to leverage expertise in scaling vehicle production.
DoorDash acquired Metis last year to infuse AI-native thinking into the organization. The company has seen significant AI spend increases, with June expenditures rising 20x compared to January levels, though spending has since flatlined through intentional optimization efforts.
Dashbench benchmark was released to evaluate model and harness performance on coding tasks, enabling ROI calculations on AI investments. The majority of AI spend remains within engineering tasks, though non-technical organizations show the highest growth in AI tool adoption.
Frontier lab collaborations revealed gaps between model performance on simplified tasks versus enterprise data scenarios. Models that perform adequately on scrubbed data may underperform when applied to real organizational data with actual complexity.
The challenge involves determining whether performance gaps result from harness configuration needs or inherent model limitations in data distribution or capability sets for accounting, analytics, and finance functions.
Despite robotics and AI advancement, DoorDash predicts more Dashers will be needed in 10 years rather than fewer. With 9 million current Dashers and 25% year-over-year business growth, scaling to 5x or 10x current operations would require finding supply from sources beyond human capacity alone.
The multimodal approach combining robotics, drones, and AI will complement rather than replace human Dashers, enabling DoorDash to meet growing demand across expanded service areas.
DoorDash envisions operating a multimodal fleet that incorporates every possible delivery modality available. This includes increased deployment of autonomy and robotics alongside sustained or even expanded human involvement. The introduction of autonomy and robotics is expected to generate significant efficiency gains over time, which in turn will drive stronger demand as delivery becomes more affordable.
Andy Fang identifies a notable trend where Google search queries have become longer over the past couple of years. This shift suggests that people are becoming more comfortable interacting with agents and apps in ways similar to human conversation. Looking ahead, reducing friction in how users interface with apps and agents will make it easier for people to place orders across categories including food, groceries, and retail.
Fang emphasizes the need to consider what an agent-first experience looks like. DoorDash has begun testing this approach with the DoorDash CLI launched the previous week. This shift toward more natural agent-first interactions is expected to unlock emerging use cases that would not be feasible in today's world.
A concrete example involves a startup office manager who initially planned to use the DoorDash CLI to streamline lunch orders. Upon discovering that DoorDash handles convenience items and groceries, the user began pointing a camera at their pantry shelf. When the shelf appeared empty, they would activate the agent to automatically restock items. This type of automated restocking use case becomes more practical as systems shift toward agent-first design.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
