Decoding Zoox: Building Autonomous Ride-Hailing Service

June 30, 2025.

It’s not often Amazon knocks on your door to offer $1B to acquire your company. But when it does, make no mistake, it’s because it has a grander vision for your product. The same happened in the case of Zoox.

In 2020, Zoox, a company pioneering autonomous ride-hailing with its custom-built, pill-shaped robotaxis was acquired by Amazon. Over the last five years, the team at Zoox has rigorously tested its service across seven major U.S. cities, including San Francisco, Las Vegas, Miami, Austin, and more. With a fleet of about 200 vehicles in the U.S. as of October last year, and now with the backing of the e-commerce giant, Zoox plans to produce around 10,000 vehicles annually at its newly launched production plant in California. That’s right, we’re talking 10,000 autonomous vehicles a year.

While 10,000 might not seem like a big number, especially compared to the more than a million Teslas and thousands of Waymos operating on the streets today, it’s hard to ignore that Amazon, the primary owner of Zoox, will likely advance this technology beyond personal transportation to deliver goods and services and is evident in Amazon’s current roadmap. 

Amazon has already started transitioning its existing delivery vans to custom-built electric vans developed by Rivian, with approximately 10,000 of these vehicles now on U.S. streets. The pace at which Amazon is electrifying its fleet, combined with the production target of 10,000 vehicles per year set by Amazon for Zoox, gives us the potential and scale at which Zoox will operate in the coming years. Given this scale, it's prudent to understand the underlying self driving system developed by Zoox. 

Before we dive deep into the system architecture of Zoox, we'll first try to understand the vision of Zoox.

Zoox's Vision:  Safer & Cleaner Urban Transportation System

Zoox started with a mission to make personal transportation safer, cleaner, and more efficient for everyone by introducing a new model of mobility, a bidirectional electric car that is fully autonomous. Unlike traditional vehicles adapted for autonomy, Zoox’s goal has always been to build a fully autonomous vehicle from ground up with no steering wheel, no pedals, purpose-built to deliver safe and efficient urban mobility solutions. You might ask how? 

Well, take a look at the graphic below. Notice how those four wheels, located on all four corners, can steer independently. Theoretically, Zoox can turn in any direction while maintaining its center of gravity. In other words, it can rotate in place. That’s gold for vehicles navigating tight urban spaces.

Now that we understand the mission of Zoox and we can picture what it looks like today, you might be wondering: how did this project even begin? And more importantly, who are the creators behind Zoox?

Evolution of Zoox

Zoox was born from the bold vision of two of the most inspiring co-founders—Jesse Levinson, a Stanford PhD and key contributor to the university’s pioneering self-driving car program, and Tim Kentley-Klay, an Australian designer-turned-entrepreneur known for pioneering technologies at the intersection of art and science. Tim is now building HYPR, an ambitious AI-enabled robotics venture. 

Unlike traditional players in the market, this duo didn’t want to take an existing vehicle and simply add a layer of sensor sauce on top of it. They wanted to design the vehicle from the ground up, specifically to enable autonomy in dense urban environments where space is a luxury. Their answer? The Zoox VH1.

This was the first prototype, launched in 2015, and known for its bidirectional capabilities. What do we mean by that? If you look at traditional vehicles, they cannot turn in place. Why? Because the steering is connected only to the front wheels. In the case of Zoox, that’s not the case. All four wheels can be steered independently, allowing the vehicle to move in any direction while maintaining its center of gravity.

This vehicle is also custom-designed for Level 5 autonomy, eliminating the need for a steering wheel or a dedicated driver. While developing a working prototype is the first step in the automotive industry, making that system safe, reliable, robust, and efficient is even more critical. That’s where industrial design comes into play.

After the launch of this first prototype–VH1, the team decided to develop an industrial design, which they launched in 2016. This version later evolved into a commercial-ready prototype that was launched in 2018. It's this version that is referred to as the Zoox robotaxi that now operates on the streets of San Francisco and Las Vegas. 

Now I know, some of you must be wondering— When will I get to see this in London, Mumbai, Singapore, Hong Kong, Paris, Turin, Istanbul, or Sydney?

Well, Zoox plans to produce around 10,000 vehicles each year, and that kind of volume definitely hints at global expansion in the coming years. While we don’t yet know which cities outside the U.S. Zoox will enter, one thing is clear: Given this scale, it's important to understand the approach adopted by the engineering team at Zoox to design this vehicle.

Zoox Platform & Design  - 8 Core Principles

If you’ve read our essay Automotive Is Now Autonomous, you probably remember that the self-driving module of these vehicles boils down to four core pillars—perception, localization, path planning, and motion control. We covered these four pillars in detail in that essay. While these pillars form the core foundation of all modern autonomous vehicles what really makes a difference is the way they are implemented. The design and engineering principles that govern their implementation is unique to each of the top players in the space. You will notice that in the approach adopted by Waymo, Tesla, and Zoox. 

And that matters.

If we don't take time to understand the underlying design and the engineering principles that are unique to each of these top players in the industry, how are we going to improve this technology? How are we going to solve problems for our customers? It’s this design and engineering approach that each of these players took in this space that makes each of these vehicles novel and valuable. We covered the approach adopted by Waymo and Tesla in the first two deep-dives we did in this series on autonomous vehicles. 

Now, you might ask—well, so what makes Zoox’s approach novel? 

There are 8 core engineering principles that Zoox has built into its system, each one tailored to deliver reliable, safe, efficient and autonomous mobility solution in dense urban neighborhoods. What are these 8 principles? We will now cover them one by one.

1. A Symmetrical Platform for Urban Navigation

Zoox’s platform is built around a unique four-quadrant design, where all four corners of the vehicle are completely symmetrical. This means that the sensors, wheels, and steering systems are identical across each corner.  It enables omnidirectional maneuverability: the vehicle can rotate in place and thread through tight urban corridors with surgical precision. You might ask, why does it matter? In modern cities dotted with millions of stoplights, dense traffic patterns, and Pokémon-Go-styled turns, this ability to maneuver with such precision isn’t a luxury—it’s a necessity. 

It’s this extreme navigational agility that ensures Zoox vehicles spend less time stuck in traffic and more time in delivering value to end users. It significantly reduces the urban inertia—the “lost in traffic” mode that engulfs most modern cabs on the roads today. Their size and design limit their ability to maneuver and, consequently, their ability to provide top-tier service to their customers. Zoox understood this problem and stood up for it from day one. Their answer? Symmetrical design married with a bidirectional architecture.

2. Bidirectional Architecture

When we discuss the architecture of traditional cars, we use the terms "front" and "back". That doesn't technically apply to Zoox vehicles because they are fully bidirectional. That means they can move forward and backward with equal ease without needing to perform a U-turn. This architecture, tightly coupled with the symmetrical design, gives Zoox the agility needed to navigate one-way streets, dead ends, and dense traffic patterns in urban environments. While being agile makes you more efficient it doesn't by default equate to safety. So what is Zoox doing to ensure these vehicles provide safe and secure service to passengers? 

3. Safety-First System Design

Any guesses? If we take a step back and look at the number of car accidents in the US alone, it boils down to roughly 6 million per year. That’s correct, that’s a huge number. In contrast, we see about 27 all-level incidents per year across more than 35 million flights in the aviation domain. Inspired by that level of mission-critical engineering and reliability, Zoox has adopted aerospace-grade principles to design a vehicle that prioritizes safety at every level. Their system integrates over 100 proprietary engineering innovations enabling multi-layered operational resiliency across the entire tech stack—perception, prediction, and motion control.

Why do we need such an intense focus on safety and vehicle resiliency? Because 94% of road accidents are caused by human error. To drive safely, a Zoox vehicle must anticipate and mitigate these failure modes. It has to understand the current state of nearby drivers and accurately predict their future actions. Every decision—from perception to trajectory prediction to motion control—is validated through layers of simulation and real-world testing. Need proof?

Zoox's Agent Trajectory Prediction System

Take a look at the image on the left side in this graphic. It’s an intersection filled with a variety of vehicles and pedestrians engaged in diverse activities. If we try to decode this scene, we realize that some of those folks are staring at their phones, some are enjoying a uni-soul city walk, and others are enjoying their time with their family. 

Now in order for a Zoox vehicle at the intersection to make accurate decisions around how it would navigate such an intersection, it needs to predict the future state of all of those agents with the highest levels of precision. That’s the only way to ensure it doesn’t run into any of them—neither the nearby vehicles nor pedestrians. So how does it do that? Well, first it splits the incoming feeds into a series of semantic layers. A semantic layer could represent pedestrians, bicycles, nearby surroundings, etc. Basically, we are trying to ensure that we analyze each of those types of agents present in the scene independently. That’s what you see in the image on the right. Each layer in the image on the right hand side of that graphic represents a specific type of agent and their position. 

Now that we know the type of agents and their current positions, our job here is to make accurate predictions about where they will be 5–10 seconds from now. If we can predict that accurately, then our vehicle can make safe and reliable moves on the road. Confused?

Take a look at this graphic below. This is a truck navigating a 3-way intersection. The green boxes represent the potential positions identified by the Zoox software system—showing where the agent could be up to 6 seconds into the future. The blue box represents where the agent actually went. Each path is a possible future state of an agent as generated by the prediction system, with an associated probability for that specific agent in the scene.

Now that we know these potential future states for one agent, we should be able to predict the same for all agents in the scene. If you look at this graphic, you’ll see that Zoox can do exactly that for hundreds of agents in real time.  Here, you see an example of a Zoox vehicle negotiating a busy intersection in Las Vegas at night. The green boxes show the most likely prediction for each agent in the scene—extending as far as 8 seconds into the future.

You might ask, even with this level of engineering resiliency, what if it still gets into an accident? Zoox doesn’t have a front or rear buffer like traditional vehicles, so how does it protect passengers in a crash? How do you make it crash-safe?

4. Crash Energy Dissipation

You engineer it to absorb crash energy from all sides. Zoox wraps the passenger cabin in a safety shell, much like a protective core, ensuring that if something does go wrong, the corresponding impact energy gets absorbed and not transferred to passengers. How does it do that? 

The Zoox vehicle features a fully encapsulated passenger safety cell—a rigid internal core designed to protect passengers from multi-directional impacts. This cell is designed to withstand structural deformations commonly seen during an impact. Surrounding this core are energy-absorbing modules engineered to deform progressively during a crash, diffusing kinetic energy released during the crash uniformly across the frame to minimize the impact on the passengers. 

This structure is made of high-strength materials and layered composites that distribute crash forces uniformly across the frame while preserving cabin integrity. This approach enables robust crash dynamics in edge-case scenarios where impact vectors are in most cases unpredictable. But even with this level of crash-resilient engineering and top-tier prediction and simulation models, what happens when a Zoox vehicle encounters a scenario that it has never experienced before and has not been trained to handle? How does it respond to that scenario? That’s where the next layer comes in: human-in-the-loop safety via tele-guidance.

5. Tele-Guidance for Edge Cases

No matter how smart a robotaxi is, the real world is messy.  You start testing a self-driving system and soon run into what’s known as the long tail problem—millions of unpredictable, low-frequency yet high-risk scenarios that your vehicle may not be trained to handle like construction zones, blocked lanes, unmarked intersections, or erratic pedestrians.

To handle these scenarios, Zoox has built a tele-guidance system into its core architecture. When the vehicle detects a low-confidence or ambiguous scenario, it pauses and sends a real-time assistance request to a centralized remote operations team also known as the remote mission control team. This team monitors the high-fidelity sensor data and live feeds streamed by the vehicle and uses a custom interface to drop digital waypoints—safe, high-level navigation cues that the vehicle can follow autonomously.

These waypoints are interpreted and executed by the vehicle's autonomous stack. In other words, humans intervene at the intent level, while the vehicle retains control over execution and safety constraints. The result? Continuous autonomous operation with minimal human interruption. Now that we’ve covered Zoox’s approach to safety and adaptability, you might ask—what about passenger comfort? What about the in-cabin experience itself? What does it actually feel like to ride in a Zoox?

6. A Passenger-Centric Experience

Well, if you ask the riders in California and Seattle, they’re giving it a thumbs up. No driver seat. No steering wheel. That opens up space—and Zoox puts it to good use. The cabin is designed around the passenger experience, not the driver. Think: wireless charging near every seat, personalized temperature control system, and immersive sound systems. It’s like a rolling lounge for your next commute. You might ask, what about engines and the fuel? 

7. All-Electric Dual Battery System

Zoox runs on a dual-battery electric powertrain built for urban duty cycles. One charge powers it through a full day of operation. It’s clean, efficient, and reliable—three boxes that every robotaxi has to check. In simple words, it significantly reduces the operational cost per mile for fleet operators. Now, some of you might ask: Given we’re soon going to see 10,000 of these on the streets, what about testing in real-world conditions?

How does Zoox ensure these vehicles and the corresponding AV system is proven outside the lab?

8. Real-World Testing with Production Fleet

Before Zoox deploys its custom vehicle on public roads, it first trains the system using a fleet of Toyota Highlanders retrofitted with the same sensor stack found on Zoox’s production vehicles. This test fleet allows Zoox to collect large-scale data from newly targeted urban environments where service is about to launch. That data includes road conditions, traffic patterns, weather patterns, pedestrian behavior, and local driving styles. All of it is streamed back into Zoox’s centralized autonomy stack—or more precisely, into the neural networks that govern perception, prediction, and planning algorithms. 

Once the system has been trained and validated, that updated version of the autonomy stack is deployed to the production fleet that will operate in the given neighborhood. This process allows Zoox to surgically fine-tune its autonomy stack to each new urban environment that it intends to launch , enabling safe and reliable service.  That leads us to an important question: What does the autonomy stack of Zoox actually look like? And how does Zoox’s engineering approach across its autonomy modules differ from those adopted by Tesla and Waymo?

While we have covered some of the engineering principles adopted by Zoox in its autonomous stack in today's essay, we cover this system in great detail in our Self-Driving Cars course. In this course, you'll learn how to build a self-driving system from the ground up. You'll gain a deep understanding of the autonomous stack developed by Tesla, Waymo, NVIDIA, and Zoox. Our courses have helped over 80,000 students from 150 countries kick-start their careers in these emerging fields.

If you're serious about entering the self-driving car industry, this course is for you. 

Cheers,

Prathamesh

Disclaimer: This blog is for educational purposes only and does not constitute financial, business, or legal advice. The experiences shared are based on past events. All opinions expressed are those of the author and do not represent the views of any mentioned companies. Readers are solely responsible for conducting their own due diligence and should seek professional legal or financial advice tailored to their specific circumstances. The author and publisher make no representations or warranties regarding the accuracy of the content and expressly disclaim any liability for decisions made or actions taken based on this blog.