Nvidia's Gives Away for Free Open Source Robotic Driving Model --- The End of the Driver Era Is Here: Robotaxis Arrive Even in Europe

 

"A Chinese giant is challenging Americans on the streets of London. Autonomous taxis have ceased to be merely a futuristic vision and have become the battleground for a fierce fight for billions of dollars. When will they arrive in Poland?

 

By the end of 2025, the global fleet of fully autonomous taxis – operating on roads without a driver – exceeded 12,000 units.

 

Although in terms of revenue and operational sophistication, the American company Waymo leads the world, in terms of fleet size, China dominates, with Baidu at the forefront (Apollo Go service). According to market analyses by TrendForce and announcements from operators, the coming year 2026 is expected to bring a rapid acceleration of growth – the global number of robotaxis could reach as many as 35,000 vehicles. But already in 2030 – as projected by Goldman Sachs and UBS – there are expected to be 3 to 5 million such vehicles.

A Business That Has Started to Make Money

 

"Robotaxis have reached a turning point, both in China and the US," says Baidu CEO Robin Li, quoted by CNBC. His words are strongly supported by the numbers: in the third quarter of 2025, Baidu's Apollo Go fleet completed as many as 3.1 million rides, and the service is becoming commonplace in major metropolises. Li thus emphasizes that the technology has moved beyond the experimental phase.

 

What was "burning cash" just a year ago is now becoming a profitable model. Baidu announced that it has achieved profitability in Wuhan and other Chinese cities. This has happened despite the fact that fares remain competitive (often 30% lower than traditional taxis). The scale of operations – exceeding 250,000 fully autonomous rides per week – allows for cost balancing, demonstrating that a business model based on autonomy can be scalable even in markets with lower purchasing power. “If we can generate a profit on one car in a second-tier city in China, we can generate profits in many cities around the world,” said Halton Niu, CEO of Apollo Go's international operations. “Scale matters. If you only deploy 100 to 200 cars in one city, you will never become profitable,” he concludes.

 

Robotaxi operations are benefiting from falling costs of key components such as lidar and batteries, and improved safety.

 

The cost of producing an autonomous vehicle (Apollo RT6) has fallen to approximately $28,000 – a price comparable to a well-equipped mid-range passenger car.

 

Chinese companies are expanding abroad, targeting the Middle East and Europe. Apollo Go obtained further permits in Abu Dhabi this year and also established a strategic partnership in Switzerland. Pony.ai, in turn, plans to launch a fully autonomous commercial robotaxi business in Dubai in 2026. WeRide is also expanding its presence in the Persian Gulf.

 

This direction of expansion is not accidental – Chinese companies are avoiding the US market due to technological restrictions imposed in 2025 on Chinese software in vehicles.

 

As a result, Chinese robotaxis are increasingly relying less on American Nvidia processors, switching to domestic chips from Huawei (Ascend series) or Horizon Robotics. Although Chinese computing power for AI training is still less efficient than Tesla's or Waymo's clusters, China is making up for it with infrastructure. In cities like Wuhan, a "smart road" communicates with the car, taking some of the computational burden off the vehicle's processors. This is a development model that Americans – who are not investing in road infrastructure – cannot easily implement.

 

However, the American giant is not giving up. In response to growing competition and market fragmentation, Nvidia made a move in mid-December 2025 that could turn the tables. The company released its latest AI model for autonomous driving – DRIVE Alpamayo-R1 – as open source. This solution utilizes so-called "Physical AI" and "chain-of-thought reasoning." Thanks to this, vehicles not only "see" obstacles but can logically analyze complex and unusual traffic situations – for example, unusual pedestrian behavior – in a way similar to the human brain [1].

 

The Battle for Big Cities: Landings in London and Austin

 

While China is storming foreign markets, American rivals have so far focused on their own backyard. In the US, Alphabet's Waymo has outclassed the competition, reaching 450,000 paid rides per week in December 2025. The company's fleet has grown to over 2,500 vehicles, and the service is already available in Phoenix and San Francisco, as well as in new cities: Austin and Atlanta. Now, however, the company is also preparing for foreign expansion. Next year, it wants to enter London.

 

However, Baidu with Apollo Go beat them to it, which caused quite a stir in the industry. According to the BBC, the Chinese have officially announced their entry into the British capital, utilizing the new legal framework of the Automated Vehicles Act. This law was enacted with the aim of making Great Britain a leader in autonomous vehicle technology. Baidu, a leader in autonomous vehicles, sees London not only as an operational testbed for left-hand traffic conditions but, more importantly, as a prestigious gateway to the rest of Europe.

 

The first test units, in collaboration with Uber, are expected to appear on London streets in the first half of 2026. Initially, the fleet will likely consist of 20 to 50 vehicles, which will focus on mapping the area and calibrating systems for the specific bottlenecks of London's City and Canary Wharf. Ultimately, Baidu plans to increase the number of vehicles to several hundred, allowing them to compete effectively with traditional black cabs.

 

 The latest Apollo RT6 model is expected to be deployed in London – a vehicle designed from the ground up without a steering wheel, intended to demonstrate a technological advantage over Waymo, which still relies on modified production cars like the Jaguar I-Pace.

 

Meanwhile, in the USA, Tesla has launched its own offensive. Unmarked Model Ys have appeared on the streets of Austin, Texas, performing fully autonomous rides – without a safety driver behind the wheel. Elon Musk officially confirmed: "Tests are underway without passengers inside." This is a breakthrough that investors have been waiting for for years – the company's shares reacted with an immediate increase, and Tesla proved that its camera-only system is ready to compete with the lidar-based systems of Waymo and the Chinese companies.

 

Although production of the dedicated Cybercab robotaxi has been confirmed for April 2026, the tests in Austin show that Musk does not intend to wait for the new platform but wants to win the race using the current fleet. This is a risky strategy, betting everything on artificial intelligence, while the rest of the industry still relies on precision lasers.

 

 In the US, Tesla and Waymo are gaining new competitors. Zoox (Amazon), which opened a factory in Hayward this year and obtained key regulatory permits, is preparing for this fierce competition.

 

Poland is in the waiting room. When will robotaxis appear in Poland?

 

While the revolution is accelerating in the US and China, Poland must be patient. In Poland, fully autonomous taxis – ordered via an app and arriving without a driver – will probably not be available until 2028-2030. This is because Poland, as part of the EU market, is subject to restrictive regulations from Brussels, which is only now finalizing the legal framework for mass autonomous traffic. Unlike the liberal approach in Texas, California, or centrally controlled China, Europe is prioritizing caution, which delays the debut of giants like Waymo on the Old Continent. In 2025, the key issue became not only safety but also data sovereignty – the EU fears the transmission of data from Chinese car cameras to servers outside Europe.

Poland, as part of the EU market, is subject to restrictive regulations from Brussels, which is only now finalizing the legal framework for mass autonomous traffic.

 

The barrier is not only legal but also geographical. What works in sunny Phoenix or Dubai will not necessarily work in Poland. Polish winters, with slush covering the lanes and freezing rain, pose a challenge for current sensors. Therefore, robotaxi operators are primarily choosing markets with milder climates. However, this does not mean that nothing is happening. Instead of American taxis, the pioneers in Poland may turn out to be autonomous minibuses from the Polish company Blees from Gliwice, which is conducting tests of its vehicles in Silesian cities. The company's CEO, Tomasz Pacan, estimates that due to current regulations, there must be an operator on board to ensure passenger safety and who can take over manual control or apply the emergency brake at any time. "All this is so that passengers can feel safe," said CEO Pacan.

 

It is also possible that the Croatian project Verne, created by Mate Rimac, which plans to expand into Central Europe after 2026, will enter Polish streets sooner than Waymo.

 

For Polish passengers, robotaxis will most likely not appear as a separate new brand, but as an additional "driverless" option in already familiar applications such as Uber, Bolt, or FreeNow.

The fight for safety and billions

 

Safety remains crucial for regulatory approval. None of the six main operators in the world have yet reported any fatalities caused by the system. Waymo boasted a 70% reduction in accidents resulting in injuries compared to human drivers. These are hard data that are beginning to convince regulators, although challenges remain, as shown by the recent power outage in San Francisco – robotaxis suddenly stopped in the streets, blocking traffic.

 

Despite the challenges, the stakes are enormous. This is a global market with significant growth potential, worth over $25 billion by 2030, according to Goldman Sachs estimates. Strategic alliances are proving to be the key to unlocking this treasure trove. The partnerships between Baidu, Pony.ai, and WeRide with Uber began to bear fruit in 2025, allowing the companies to develop robotic vehicles. This allows them to operate more efficiently. Instead of building their own user base from scratch in new markets, the Chinese are entering an existing ecosystem, which significantly reduces customer acquisition costs.

 

However, this breakthrough extends far beyond the taxi market. The development of robotaxi fleets in 2025 has become a testing ground for the future of personal cars. If the algorithms prove successful in public services, they will pave the way to a world where car ownership no longer requires driving.

 

This is a vision of the future where the "driver" becomes a relaxed passenger in their own car, no longer having to struggle with traffic jams or long journeys.”

 

1. What does development environment for NVIDIA DRIVE require in a compatible Windows or, most commonly, a Linux system, what is the necessary NVIDIA hardware, and what are the prices for these resources to be able for NVIDIA DRIVE to do refinement learning from videos?

 

The NVIDIA DRIVE platform (including DRIVE OS, DriveWorks SDK, and related tools) is primarily designed for autonomous vehicle development. Access to the full SDK and downloads requires membership in the NVIDIA DRIVE AGX SDK Developer Program (or equivalent, often for approved companies, researchers, or institutions via NVIDIA Developer or NVONLINE portal).

Development Environment Requirements

The setup involves a host machine (for development, cross-compilation, and flashing) and a target hardware (DRIVE AGX Developer Kit).

 

Host System (most common): Ubuntu Linux (typically 18.04 or 20.04 LTS, x86_64). This is the standard and recommended environment. NVIDIA SDK Manager is used for installation on Ubuntu hosts.

Windows Support: Limited or not directly supported for the full DRIVE development flow. Some tools like Nsight Systems may run on Windows, but core DRIVE OS/DriveWorks installation and flashing require a Linux host. No official Windows host is mentioned in current documentation.

Additional Host Needs: NVIDIA GPU recommended for running samples (e.g., GeForce/Quadro/Tesla with appropriate drivers), internet access for SDK Manager/NGC downloads, and Docker for some workflows.

Target System: Runs DRIVE OS (Linux-based, with safety extensions) on the DRIVE AGX hardware.

 

Installation is typically via NVIDIA SDK Manager on the host, which downloads and flashes DRIVE OS/DriveWorks to the target.

Necessary NVIDIA Hardware

The core hardware for development and runtime is a DRIVE AGX Developer Kit:

 

Current mainstream: DRIVE AGX Orin Developer Kit (up to ~254-275 TOPS AI performance).

Next-gen: DRIVE AGX Thor Developer Kit (higher performance, up to 1000+ TOPS sparse, Blackwell-based).

These kits include the SoC, reference board, and interfaces for sensors (cameras, LIDAR, radar, etc.).

For production deployment, custom boards based on Orin/Thor SoCs are used, but development starts with the dev kit.

 

The platform supports sensor suites via DRIVE Hyperion reference architecture (e.g., multiple cameras, radars, ultrasonics, LIDAR).

Refinement Learning from Videos (Fine-Tuning/Training)

"Refinement learning from videos" likely refers to fine-tuning or training deep neural networks (DNNs) for perception, prediction, or other AV tasks using video data (e.g., camera feeds).

 

On the DRIVE AGX Hardware: The dev kit supports inference and some lightweight training/fine-tuning via CUDA, TensorRT, cuDNN, and DriveWorks modules. However, heavy training (especially from large video datasets) is compute-intensive and typically done off-device.

Recommended for Training/Refinement: Use NVIDIA DGX systems (data center GPUs, e.g., DGX A100/H100/B200) or cloud instances for efficient training on video datasets. Tools like NVIDIA TAO Toolkit (Transfer Learning Toolkit) allow fine-tuning pre-trained models with custom video data. Video processing can leverage libraries like NVVL for accelerated loading/decoding.

DRIVE ecosystem integrates with simulation (DRIVE Sim/Omniverse) for synthetic data generation to augment real videos.

 

The DRIVE AGX kit itself can run refined models post-training but is not optimized for large-scale video-based training/refinement loops.

Prices

Prices are not publicly listed in detail (often require quotes or program approval), but available information:

 

DRIVE AGX Orin Developer Kit: Not publicly priced; available via authorized distributors (e.g., Arrow Electronics) for approved buyers. Requires "Get a Quote" and program membership. Historical/related kits (e.g., older Xavier/Pegasus) were in the high thousands to tens of thousands USD.

DRIVE AGX Thor Developer Kit: Recently available for preorder/general access; pricing similarly via quote/distributors (no public MSRP found).

Related note: The non-automotive Jetson AGX Orin Developer Kit (similar but not DRIVE-grade) is around $1,000–$2,000 depending on config, but it's not compatible for full DRIVE AV development.

Training hardware (e.g., DGX systems): Start from tens to hundreds of thousands USD, depending on model.

 

For exact pricing and purchase, contact NVIDIA or authorized partners through the Developer Program, as these are enterprise/automotive-grade tools, not consumer products.