The "Three Gates" of Robotaxi

Model Y is very busy.

In China, it is a "stayer" on the PPTs of major automakers' press conferences and is often used for comparison. Being "crushed and surpassed" is its fate. However, on the other side of the ocean, Model Y presents a completely different image and has entered a Next Level that few other automakers can reach.

At the end of June, an unmodified production Model Y set off from the Austin factory, autonomously drove on urban roads and highways, and finally accurately arrived at the customer's doorstep for delivery - the whole journey was 116 kilometers without any human intervention.

During the same period, Tesla's Robotaxi service opened in a small area in southern Austin, starting to carry paying passengers to experience driverless travel.

Even earlier, in Tesla's US factory, newly produced cars could automatically drive into the logistics area, eliminating the need for parking drivers.

The above events indicate that Tesla is really starting to develop towards the literal meaning of FSD (Full Self-Driving).

Tech fans and Tesla enthusiasts are overjoyed, and the cheers about "the future is here" and "looking forward to effortless victory" are getting louder and louder. People are looking forward to whether the pie drawn by Musk N years ago - allowing private Teslas to run as Robotaxis to make money when idle - is really within reach?

However, the industry is more calm.

Compared with Waymo, Baidu Apollo Go, and Pony.ai, which have been deeply involved in the Robotaxi field for many years, Tesla's current trial operation scale is extremely small (small fleet, limited area, and short mileage). It is more like a public test than an operation. Its charge of $4.2 per ride is more like an experience price, not even comparable to an "opening promotion".

Of course, it cannot be denied that the global influence of Musk and Tesla has brought high attention back to Robotaxi. However, for Robotaxi to achieve large-scale implementation and profitability, it still needs to break through the three major mountains of technical safety, cost control, and responsibility definition step by step. There is no room for empty talk or "skipping levels".

01

Technical Route Divergence: Mass Fleet VS Elite Fleet?

Tesla's Robotaxi directly uses the existing hardware (i.e., the pure vision solution) and the FSD system that is the same as the production version without any additional modification or adjustment.

This is a completely different technical route from mainstream L4 autonomous driving companies, with an obvious Tesla style - radical and bold.

This approach makes it possible for ordinary car owners' private cars to seamlessly connect to the Robotaxi network to "take orders and make money". However, objectively speaking, the "foundation" of this system is still L2-level assisted driving.

The original intention of L2 is to improve the driving comfort of human drivers, emphasizing the "shared" driving between humans and machines and relying on drivers to support the system. In contrast, L4 requires the system to be safe and autonomous, with safety as the top priority and the goal of "zero accidents". The two are fundamentally different.

What Tesla is currently doing is actually using the "foundation" of L2 level to perform the "tasks" of L4 level - allowing the vehicle to carry passengers autonomously without a driver.

As for the safety redundancy particularly emphasized in L4 level, Tesla's answer is "temporarily rely on humans".

Currently, there is a safety operator sitting in the front passenger seat of Tesla's Robotaxi. According to passengers who were invited to experience it, during the vehicle's operation, the safety operator's right hand was always placed on the front passenger's threshold, presumably the location of the emergency button. In a sense, the safety operator acts as a temporary "perception and safety redundancy".

The safety operator is ready to make an emergency manual brake at any time

Obviously, this is not a long-term solution. True driverless driving must eliminate the safety operator.

It is still unclear whether the steering-wheel-less driverless taxi Cybercab released last year will be equipped with hardware redundancy. However, it is certain that whether it is the current Model Y running as a Robotaxi or the future Cybercab, Tesla adheres to achieving autonomous driving through cameras and machine learning.

Musk seems to be betting that the combination of his pure vision algorithm, massive data, super computing power, and software iteration can solve all known and unknown problems. His famous quote "Humans can drive without emitting lasers" is the core of his concept.

However, just as human drivers can be blinded by oncoming high beams, confused by strong reflections on the road during a sunshower, or unable to see the road clearly in bad weather such as rain and fog, the pure vision solution that mimics the human eye can also lead to wrong decision-making outputs due to wrong perception inputs under strong light interference and bad weather.

In L2 assisted driving, these "errors" can be corrected manually. However, L4 autonomous driving cannot tolerate or accept any minor "errors".

During the two-week small-scale test so far, Tesla's Robotaxi has exposed many problems: the vehicle once drove into the oncoming lane, braked suddenly without warning, swerved while speeding, signaled a turn but actually went straight, and even parked illegally at an intersection.

Just a few days ago, Tesla's Robotaxi had its first collision accident. It suddenly turned while stationary and collided with a car parked by the roadside. These various situations all indicate the potential risks of the pure vision solution in real and complex environments.

Tesla's ROBOTAXI's first collision

The ultimate goal of L4 Robotaxi is to be "safer than humans", not "the same as humans". Whether the "mimicking the human eye" solution can achieve the goal of "surpassing the human eye", Tesla needs a larger fleet size, a wider operating range, and more driving mileage to prove itself.

If Tesla's Robotaxi wants to build a "mass fleet", then the industry mainstream represented by Waymo, Baidu Apollo Go, and Pony.ai has chosen a more stable but more costly "elite fleet" model.

The core of this model lies in "hardware redundancy" and "multi-sensor fusion", piling up sensors regardless of cost to ensure foolproof safety.

Waymo's sixth-generation sensor suite has 13 cameras, 6 radars, and 4 lidars, and the vehicle is also equipped with a series of external audio receivers. Compared with the previous generation, the number of sensors has increased, the resolution and accuracy have been improved, and the distribution positions have been optimized to ensure that there are no blind spots in the vehicle's field of vision [1].

Pony.ai's latest seventh-generation Robotaxi is equipped with 9 lidars, 14 cameras, 4 radars, 4 microphones, 2 water sensors, and 1 set of collision sensors, achieving 360° non-blind spot detection of objects and the environment within a range of 650 meters around the vehicle.

The purpose of "arming the vehicle to the teeth" is very clear. Through the cross-verification of various types and a large number of sensors, it ensures accurate environmental perception under any lighting environment and any weather conditions. Relying on 100% automotive-grade hardware and redundant design, the system can still operate safely even if a certain component fails.

If we understand it from Musk's "anthropomorphic" perspective, multi-modal sensor fusion may be like the integration of multiple human senses (vision, hearing, touch, etc.) to form an accurate and comprehensive perception of things. A common example in daily life is that myopic people often use the double confirmation of "hearing the voice and seeing the lip movements" to clearly understand what the other person is saying.

This "high-price, high-performance" route creates an "elite chariot" specifically designed for autonomous driving operations.

Of course, there is no superiority or inferiority, let alone right or wrong, in technical routes. Whether it is the "mass pure vision" or the "elite multi-fusion" route, they ultimately need to be measured by the two hard indicators of "safety" and "stability".

Similar to the gradual development of L2 assisted driving (starting from parking assistance, then highway driving, then urban driving, and finally full-area driving), L4 autonomous driving/Robotaxi also has a more strict and detailed path to "prove its strength":

From intensive testing on closed roads, to demonstration applications and operations on limited public roads, and then to commercial operations within an electronic fence;

From having a safety operator in the driver's seat, to having a safety operator in the front passenger's seat, and then to having a remote safety operator;

From free during the testing phase, to discounted during the demonstration phase, and then to market-based pricing during the operation phase.

Only by meeting the requirements of the previous stage (total mileage, operation duration, number of vehicles, etc.) can one enter the next stage. Whether it is Baidu Apollo Go and Pony.ai in China or Waymo abroad, they basically follow this path step by step.

Judging from this standard, Tesla's current configuration of "front passenger safety operator + small area + small number of vehicles" can only barely be considered to have entered the early demonstration application stage (and the number of vehicles is still far from enough). In contrast, mainstream L4 companies crossed this stage several years ago.

Pony.ai's CTO, Lou Tiancheng, once said bluntly: "Only companies that can achieve an unmanned operation scale of 500,000 hours are qualified to be at the table."

Measured by this hard standard, Tesla's Robotaxi has not really "sat at the table" yet.

This is similar to the "10,000-hour rule". To achieve professional success in a certain field, spending time is an inevitable path.

This is true for people learning professional skills and cars learning autonomous driving.

02

Business Closed Loop: Cost, Scale, and ROI

Proving that it can play the game with "safety" and "stability" is just the entry ticket for Robotaxi to the table.

The next more severe challenge is how to run through the business closed loop and prove that it can win on the premise of ensuring safety - winning in terms of cost, scale, and sustainability.

From the "safety table" to the "profit table", the game rules are completely different. The key to victory has shifted from absolute safety verification to a comprehensive competition in multiple dimensions such as cost control, operation efficiency, scale expansion, traffic entry, user acceptance, and policy adaptation.

The first "life-and-death hurdle" is cost control.

The cost here does not simply refer to the price of manufacturing a single vehicle (BOM cost), but a broader concept that includes: single-vehicle cost, maintenance cost, safety redundancy cost, and platform operation cost, etc.

Industry analysis data shows that although the cost structures and actual expenditures of each company vary slightly, the comprehensive cost still exceeds the current operating income. In simple terms, it is a case of spending more than earning, which has led to the widespread loss of current mainstream Robotaxi companies.

Last year, Waymo, with a weekly travel volume of over 100,000 trips, still caused its parent company Alphabet to incur a loss of nearly $2 billion.

Pony.ai's first financial report after going public showed that its net loss in 2024 reached as high as $275 million (approximately RMB 1.996 billion), more than doubling the loss in 2023 ($125.3 million).

No matter how wealthy a company is, it can't afford such a large amount of burning money. Cost control has become an urgent task for the players at the table.

Among the broad costs, "miscellaneous" expenses such as safety redundancy, daily maintenance, insurance, internet fees, energy replenishment, and ground crew labor are difficult to significantly reduce in the short term (Cybercab's automatic cleaning tool may help). Therefore, the single-vehicle manufacturing cost is currently still the main battlefield for cost reduction.

Tesla is developing an automatic cleaning robot for ROBOTAXI

The golden rule for cost reduction is nothing more than two things: reducing the unit manufacturing cost and expanding the production scale to generate economies of scale.

In terms of large-scale manufacturing, Model Y, with a global ownership of 5 million, is undoubtedly the king. The confidence for Musk to shout the slogan of "mass fleet, mass order-taking" comes from Tesla's cost synergy, simply put, "killing two birds with one stone": one set of hardware serves two purposes, eliminating the huge cost of separately developing, purchasing, and installing dedicated hardware for Robotaxi.

The R & D investment in its FSD software is also shared by the large number of production vehicle users. Every software iteration and upgrade not only serves private car owners who have purchased the FSD function but also empowers the future Robotaxi network. This leverage effect of R & D cost is difficult for companies specializing in Robotaxi to achieve.

In contrast, for L4 autonomous driving companies deeply involved in Robotaxi, the heavy "weapons" that once supported their qualification to be at the table, such as lidars and other devices, have become a cost disadvantage under the cost criterion.

The good news is that thanks to the highly competitive automotive supply chain in China, the "elite fleet", which was once thought to cost millions of yuan, now has a single-vehicle cost close to that of ordinary passenger cars.

Pony.ai's seventh-generation Robotaxi uses automotive-grade mass-produced components such as the same lidar as Li Auto and the same intelligent driving chip as NIO. The total hardware cost has dropped to 270,000 yuan, a full 730,000 yuan lower than the first generation in 2017, and it is expected to continue to reduce costs by 30 - 40% in the future, entering the "below 200,000 yuan" range;

The cost of Baidu Apollo Go's sixth-generation vehicle has decreased by 60% compared with the previous generation, and the price is just over 200,000 yuan;

The vehicle cost of Waymo's pure-electric Zeekr RT, which is currently under testing, is expected to drop below $100,000;

Tesla's Cybercab (target) is planned to be mass-produced in 2026, with the target cost controlled within $30,000.

After the cost is reduced, the second step is to expand the scale.

Currently, the fleet sizes of leading players are steadily increasing: