Robotaxis

October 7th 2025

Welcome to the first edition of our new deep dive series. Each piece will take one frontier technology, the kind that has the potential to reshape our world, and unpack it in full. The goal is simple: to move beyond headlines and ask the right questions. How does the technology actually work? Where is it today? What still needs to happen for it to be mass adopted? And who stands to benefit when it does?

We’re starting with Robotaxis because few technologies promise to reshape daily life so directly. Human-driven cars, both private and commercial, have dominated cities for over a century, but autonomous fleets promise to replace them with rides that are safer, cleaner, quieter - and cheap enough to undercut Ubers, taxis, even public transport.

Behind that promise lies a complex convergence of technologies: autonomous navigation, electric drivetrains, AI chips, and mapping systems. The economic stakes are enormous. A multi trillion-dollar transport market is being reshaped, with ripple effects across energy, infrastructure, and labor.

This deep dive will come in two parts. In today’s edition, we’ll examine how Robotaxis actually work, where deployments are happening today, what’s still holding them back, and when you might expect to see them in your own city. In Part Two, we’ll examine which companies are best positioned to capture the value of this shift - and what that might mean for investors.

If you’re new here, you can join free with one click to get more deep dives on the innovations shaping our world.

What Do We Mean by “Robotaxi”?

Before we go further, it’s worth defining terms. Not every car that can “drive itself” is a Robotaxi.

Levels of Autonomy

The industry uses the Society of Automotive Engineers’ levels of automation, ranging from Level 0 (no automation) to Level 5 (full automation):

• Level 0 (No Automation): The driver is fully in control at all times. The car might provide warnings, like lane departure alerts or blind-spot monitoring, but these are strictly driver-assist features, not automation. Almost all cars on the road today still fall into Level 0.

• Level 1 (Driver Assistance): The car can assist with either steering or acceleration/braking, but not both simultaneously. Adaptive cruise control (maintaining distance from the car ahead) or lane-keeping assist (steering support) are common Level 1 features. The driver must remain engaged and in control at all times.

• Level 2 (Partial Automation): The car can steer and accelerate, but the driver must supervise at all times. Tesla’s Autopilot and FSD are here today.

• Level 3 (Conditional Automation): The car can drive itself in certain conditions (e.g., on highways) but requires the driver to take over if asked. Mercedes Drive Pilot is the first approved example.

• Level 4 (High Automation): No driver needed but only within specific areas or conditions, called the operational design domain (ODD). Today’s Robotaxis (Waymo, Baidu, WeRide) are Level 4: they work in mapped city zones or geofenced areas but won’t pick you up outside them.

• Level 5 (Full Automation): Any road, any weather, no restrictions. This is the sci-fi vision, but experts agree it’s still years, maybe decades, away.

Why ODDs Matter

ODDs explain why Robotaxis today look so restricted. A Waymo in Phoenix might handle suburban streets with ease, but that same car couldn’t just drop into Rome or Mumbai and drive safely. Local rules, road markings, weather patterns, and even driving culture differ too much.

For Robotaxis to scale globally, they don’t need to solve everywhere driving (Level 5). They just need to handle the environments people use most - dense urban centers and highways. That’s why the current rollout is city-by-city, not nationwide.

The State of Play: Robotaxis on the Road

As of 2025, Robotaxis are no longer a science experiment. They’re on real roads, carrying real passengers, but only in a handful of places.

• United States: Robotaxi services are operating commercially in five states - California, Arizona, Texas, Georgia, and Nevada. Waymo is the clear leader, delivering over 250,000 rides per week across Phoenix, San Francisco, Los Angeles, Austin, and Atlanta. Cruise was once a major contender, but after safety incidents in 2023, California suspended its driverless permit, leaving Waymo as the frontrunner. Tesla has begun a limited pilot in Austin, but its cars are still classified as Level 2 and require driver supervision. Meanwhile, Amazon’s Zoox has launched a free Robotaxi service in Las Vegas (with paid rides planned), marking the first deployment of a purpose-built Robotaxi with no steering wheel.

• China: The most aggressive rollout is happening here. Sixteen cities already allow fully driverless rides in defined zones. Baidu’s Apollo Go has logged more than 6 million paid Robotaxi rides, and in Q2 2025 alone delivered 2.2M fully driverless trips (around 170,000 per week). Several other companies including Pony.ai and WeRide are also scaling and active in multiple cities.

• Europe & UK: Progress is slower but steady. Mercedes’ Level 3 Drive Pilot is approved for highways in Germany. Wayve, a UK startup, is testing next-gen end-to-end AI models, but no commercial Robotaxi fleets operate yet. Baidu, via separate partnerships with Lyft and Uber, is aiming to pilot services in Germany, Switzerland, and the UK in 2026, pending regulatory approvals.

• Japan & South Korea: Waymo has begun testing in Tokyo, signaling Japan as a likely early adopter outside the U.S. and China. In South Korea, SW Mobility is running free AV pilots in Seoul and there are government-backed plans pointing to broader commercialization around 2027 or later.

• Middle East: The UAE is moving fast. Abu Dhabi already hosts Uber + WeRide’s first international Robotaxi service, while Baidu aims to launch fully driverless Robotaxis in Dubai by Q1 2026 and scale toward 1,000 vehicles by 2027/28; Pony.ai targets 1,000 Robotaxis by 2028 across the Middle East.

• Singapore: The city-state is emerging as a regional AV hub. WeRide and Pony.ai have announced partnerships to launch autonomous shuttles/Robotaxis starting 2025–2026, with local operators (Grab, ComfortDelGro). WeRide has also piloted fully driverless robobus service (no safety officer) on fixed routes. Baidu has signaled plans to bring Apollo Go to Singapore and Malaysia.

So while Robotaxis are real, they’re still a novelty. Outside a handful of U.S. metros, Chinese megacities, Seoul, and Abu Dhabi, most people haven’t seen one in the wild - let alone ridden in one.

How Robotaxis Work

A Robotaxi is essentially a human driver rebuilt in silicon and steel. To understand how they work, it helps to break the system down into the same parts you’d find in a person: eyes and ears (sensors), a brain (chips and AI), memory (maps), and muscles (the car’s actuators and safety systems).

Sensors: The Eyes and Ears

Sensors are how a Robotaxi sees and hears the world. Cameras capture high-resolution images of road markings, traffic lights, pedestrians, and signs. Radar shoots out radio waves that bounce back to measure distance and speed - great in fog or heavy rain where cameras struggle. Lidar, short for “light detection and ranging,” fires millions of laser pulses per second to create a 3D map of the environment, letting the car measure shapes and depth with centimeter accuracy.

Early lidar units were once too expensive for mass adoption, but today Chinese suppliers like Hesai and RoboSense dominate production, with Hesai now the leading lidar supplier worldwide, while U.S. firm Luminar is set to supply its Iris lidar to Mercedes vehicles starting in 2026. In parallel, camera and radar suppliers - Sony, OnSemi, Samsung EM on the imaging side, and Continental, Bosch, Aptiv, and Denso for radar modules - have been scaling up their automotive businesses to meet rising demand.

The big divide is over whether lidar and radar are even necessary. Companies like Waymo, Baidu, Pony.ai, and WeRide embrace redundancy: they use all three modalities - cameras, radar, and lidar - to minimize blind spots and build trust with regulators. Tesla, on the other hand, has gone the opposite way. Elon Musk argues that humans drive with just their eyes, so cameras paired with neural networks are the only truly scalable solution. It’s cheaper, easier to mass-produce globally, and fits Tesla’s “scaling laws” philosophy that bigger datasets and more compute power naturally improve AI systems, without engineers needing to hand-craft rules. Tesla’s bet is that if you feed its networks enough driving video, the cars will eventually learn to drive anywhere, just like humans do. But it also introduces risk: camera-only systems are still less reliable in difficult conditions like glare or poor weather. Interestingly, even Waymo, the poster child for lidar-heavy fleets, is researching Tesla-style vision-only approaches with its EMMA project.

Compute: The Brain

The flood of raw data from sensors is useless without something powerful enough to process it instantly. This is the job of autonomous driving chips, or SoCs (systems on chips). These chips run neural networks that identify objects, predict their movements, and plan safe trajectories - all in real time, with near-zero latency, and without draining the vehicle’s battery.

This has created a new arms race in silicon. NVIDIA leads with its DRIVE Orin, the current workhorse of Robotaxi fleets, and its successor, DRIVE Thor, capable of 2,000 trillion operations per second (TOPS) and positioned as the backbone of next-generation autonomous vehicles. Qualcomm is pushing its Snapdragon Ride platform with BMW as a key partner. Mobileye’s EyeQ line, long embedded in consumer cars, is designed to scale into higher levels of autonomy and is being tested for Robotaxi applications. Tesla develops its own Full Self-Driving chip to power its vision-only approach, while in China, Horizon Robotics, Black Sesame, and XPeng’s in-house “Turing” chip are emerging as domestic contenders.

Maps: The Memory

Even the best vision systems struggle without context. High-definition (HD) maps give Robotaxis a memory of the road - lane-level details, curb positions, traffic light locations, even common pedestrian crossings. This allows centimeter-level localization, even if GPS drops out.

Companies tackle mapping differently. Waymo and Baidu send lidar survey cars to build precise but costly proprietary maps. Mobileye uses its REM system to crowdsource maps from millions of EyeQ-equipped consumer cars. Legacy players like HERE and TomTom are repositioning for the autonomous era, while in China NavInfo, Kuandeng, and AutoNavi play leading roles. To cut costs, firms are testing new methods such as change detection (only remapping what’s new) and “implicit maps,” where neural networks learn the patterns of roads and intersections rather than storing literal blueprints, like teaching AI the feel of a city.

The map debate mirrors the lidar debate. Waymo, Mobileye, and Baidu argue that maps are essential for safety and redundancy. Tesla scoffs, calling maps brittle and impossible to scale worldwide. Some researchers advocate a hybrid model: start with maps as training wheels, then gradually lean more on AI’s ability to generalize - that is, to handle new, unseen roads by recognizing patterns from past experience rather than relying on a pre-drawn map.

Software: The Decision-Making

Once a car can perceive its surroundings and localize itself on a map, it needs to decide how to act. Should it change lanes? Overtake? Stop for a pedestrian? This is the job of the driving software.

Most companies still use modular systems: perception models identify objects on the road, those outputs feed into prediction models that estimate what those objects will do next, and then planning software decides how the car should respond before issuing commands to the actuators. Each step is separated, tested, and monitored. This structure makes it easier to debug errors and to reassure regulators, since engineers can point to how each decision was made.

Tesla takes a different approach. With FSD v12 it has gone fully end-to-end: the neural network ingests raw video from cameras and directly outputs driving actions, skipping the hand-engineered stages. Tesla argues that this “single brain” can learn from billions of miles of data, adapt to edge cases more smoothly, and ultimately be safer once scaled. Critics counter that while end-to-end systems may prove more flexible in the long run, they are harder to interpret, harder to certify for safety, and much riskier to debug when something goes wrong.

Wayve is experimenting with a third path: vision-language-action models like LINGO-2. Unlike Tesla’s black-box neural nets, LINGO-2 can both drive and explain its choices in plain English (“I’m slowing down because there’s a cyclist ahead”). That transparency makes regulators more comfortable and could help win public trust - one of the biggest barriers to adoption.

Actuation and Safety: The Muscles and Reflexes

Finally, all of this intelligence needs to translate into physical motion. Actuators - steering, braking, throttle - are the car’s muscles. Safety systems are its reflexes. For autonomy to scale, these need to work with aviation-level redundancy.

That means dual steering systems, backup braking and power, and safety microcontrollers that monitor the main chips for faults. Many of the same suppliers used in ordinary cars - Infineon, NXP, Renesas, Texas Instruments - also provide these safety electronics for AVs. Large firms like Bosch, Continental, and ZF that deliver complete subsystems (braking, steering) directly to automakers, are adapting their platforms with redundant brake and steer-by-wire designs for automated driving.

On this front, there’s little disagreement: redundancy - having backup systems like dual steering or braking so one can take over if the other fails - is non-negotiable. The debate is mostly over how much of this backup should be built in physical hardware versus handled by software. Hardware redundancy, like separate steering racks, is safer but adds cost. Software-based redundancy is cheaper but raises questions about reliability. Waymo, Zoox, and most Chinese players lean heavily on hardware, since it’s easier to prove to regulators. Once again Tesla is the outlier relying more on software checks, consistent with its low-cost, vision-first philosophy.

What Still Needs to Be Done

Regulation: A Patchwork World

The biggest barrier to scaling Robotaxis isn’t chips or sensors - it’s laws. Every country, and in the U.S., every state, is writing its own playbook.

In the United States, there is no national framework, only a patchwork of state-level rules. Arizona, Texas, Nevada, and Georgia have been the most permissive, actively welcoming testing and deployment, while California allowed Waymo to expand but pulled Cruise’s permit in 2023 after a safety incident. Federal regulators like the NHTSA continue to experiment with exemptions and reporting requirements but have yet to impose unified rules.

China is by far the most aggressive. Although licenses are still issued city by city, national policy actively promotes pilot zones and commercial expansion, with clear political backing and guidance from the Ministry of Industry and Information Technology.

The EU’s common rules are maturing, but city-level approvals still decide timing. Germany is furthest along, targeting driverless Level-4 ride-pooling in Hamburg from 2026 under its national law; classic on-demand Robotaxis would follow case-by-case approvals. France can already authorize Level-4 services, but initial permits focus on shuttles/minibuses.

The United Kingdom has one of the clearest frameworks globally with its Automated Vehicles Act of 2024, which clarifies liability and insurance, though the secondary safety and licensing rules won’t be finalized until 2026.

In Asia, Japan and South Korea are pursuing parallel strategies. Japan is moving quickly with national approval for Level 4 operations in 40 designated areas and aims for a nationwide rollout by 2027. South Korea’s roadmap also targets around 2027 for commercialization, supported by coordinated government pilots.

Elsewhere, the Middle East is emerging as a proactive region. Dubai and Abu Dhabi have already established AV permit systems, with Dubai targeting 25 percent of all trips to be autonomous by 2030. Singapore continues to refine one of the most structured frameworks in the world, using staged pilots with shuttles before opening up to full robotaxi operations.

Australia and Canada are progressing more cautiously, each with pilots (in cities such as Melbourne, Sydney, and Toronto) but without nationwide frameworks or commercial rollouts yet.

Cost: Still Too High (except in China)

Autonomous cars today are expensive prototypes, not cheap fleet workhorses.

The biggest cost driver remains hardware. Lidar sensors have fallen from $75,000 to under $200, but each Robotaxi still needs several, plus cameras, radar, and powerful onboard computers. Chips like NVIDIA’s DRIVE Thor can cost thousands of dollars apiece, and most fleets use multiple units per car. On top of that, aviation-grade redundancy, the backup hardware required for safety certification, still adds thousands per vehicle.

Mapping adds another hidden layer of expense. High-definition maps require constant updates and re-surveying, especially in cities that change frequently. Each expansion into a new market can demand millions in up-front mapping costs before the first commercial ride takes place.

Then come operational costs - maintenance, remote monitoring, fleet management, and utilization. For Robotaxis to beat private cars on price, each vehicle must rack up tens of thousands of miles per year, running almost continuously. That requires durable components, efficient servicing, and regulatory approval for 24/7 operation.

Together, these factors keep Robotaxi fares high. In San Francisco, Waymo rides still cost around $2–$3.50 per mile - often more than Uber. However, analysts expect prices to fall toward $0.25 per mile by the 2030s as hardware gets cheaper and fleets scale, potentially making Robotaxis cheaper than private car ownership and, in many cities, even public transit. Purpose-built vehicles like Zoox’s pod, which eliminate steering wheels and driver ergonomics, should drive costs down further once mass-produced.

China already offers a glimpse of what mature economics might look like. Baidu says its Apollo Go service has reached per-vehicle profitability in some cities, helped by lower hardware/mapping costs, local suppliers, and strong policy support. For example, its RT6 purpose-built Robotaxi is priced around $34,000, with some reports citing figures under $30,000, whereas industry estimates put a Waymo-equipped Jaguar I-PACE well into six figures once sensors and compute are included. In Wuhan, reported fares have run as low as $0.31 per mile, underscoring a faster path to sustainable unit economics than most Western peers.

Safety & Edge Cases: The Decisive Metric

No matter how cheap they get, Robotaxis won’t scale unless people believe they’re safe. The bar isn’t just matching human safety - it’s being safer, consistently, and proving it.

On that score, Waymo has the strongest record. By June 2025, Waymo vehicles had logged 96 million fully driverless miles. Peer-reviewed research published in May 2025 found that Waymo cars had nearly the same overall incident rate per mile as human drivers, but were significantly safer in key categories: fewer pedestrian crashes, fewer intersection collisions, and fewer property-damage accidents. An insurance study from Swiss Re backed this up, reporting 88% fewer property damage claims and 92% fewer bodily injury claims compared to human-driven cars. Baidu doesn’t publish peer-reviewed safety studies like Waymo, but incident rates are manageable enough for regulators to keep expanding service zones.

Tesla, by contrast, is still officially at Level 2. Its Autopilot and Full Self-Driving (FSD) features require driver supervision, and regulators treat them as advanced driver-assist systems rather than autonomy. NHTSA data from 2024 linked Autopilot to 956 crashes and 13 fatalities. Many involved tricky situations where cameras struggled, such as glare from the sun, heavy rain, or stationary vehicles on highways. There have also been a few incidents during their Austin ‘Robotaxi’ roll-out. Tesla’s argument is that this record will improve quickly as its system scales: more cars on the road mean more data, and more data means smarter neural networks.

The problem is edge cases: construction zones, police officers waving cars through, unusual weather, children darting into the street. These rare scenarios are difficult to anticipate and code for, which is why most Robotaxi operators still keep remote monitors on standby to assist if the system gets confused. The goal is to push the assist ratio from early numbers (e.g., 1 remote operator per 10–20 cars) toward 1:100+ as software improves.

Bringing It All Together

The cost and safety bottlenecks are being chipped away by rapid progress. Sensors are cheaper, safety data is improving, and fleet operations are maturing. Ultimately, these issues aren’t unsolvable. The deciding factor will be regulation: not whether Robotaxis can work, but when lawmakers in each city, country, or region decide they’re ready to allow them. That’s what will determine whether you can hail a Robotaxi in Phoenix in 2025, London in 2027, or Mumbai in 2032. The technology may be global, but the rollout will always be local.

Next Time

Now you know how Robotaxis work, where they are operating right now, what’s slowing down their adoption, and when they may be in a city near you - that’s only half the story, though. In Part Two, we’ll look at who’s actually positioned to win when the world goes driverless - from the chipmakers powering autonomy to the fleets deploying it at scale.

If you’d like to get a head start, Nanalyze has already explored the investment side in depth. Their research digs into which Robotaxi companies are genuinely investable - and which promising stories don’t hold up under the numbers. You can find their analyses here.

Not a subscriber and want part two and future deep dives in your inbox? Join free with one click.

Your Thoughts

This is the first instalment of our new deep-dive series, and I’d love to hear what you think so far.

Was this the right kind of content? Anything you’d improve or want to see done differently as we build the rest of the series?

And if there are any topics you’d love to see unpacked in future editions, let me know - I’m always open to ideas.

Just hit reply - I read every message, and your feedback will help shape how these deep dives evolve.

Share It Forward

If you made it this far, you probably got something out of it - maybe a new perspective, a few insights, or even just a bit of optimism about where the world is heading.

If that’s the case, I’d really appreciate it if you shared this email with one person who’d enjoy it too.

That single act helps us reach more curious, hopeful people and keeps this project growing.

See you soon,

Max

P.S. If a friend forwarded this, you can join free with one click to get part two and future deep dives.