The robotaxi industry has been betting heavily on simulations and computer models to solve autonomous driving, but a top AI executive says that strategy is fundamentally flawed. According to Pony AI's Chief Technology Officer, the companies that win the robotaxi race will be those that collect massive amounts of real-world traffic data and learn from everyday chaos on actual streets, not those with the best lab results.

Why Simulations Alone Can't Solve the Robotaxi Problem?

Dr. Tiancheng Lou, founder and CTO of Pony AI, recently published an opinion piece arguing that the industry has overestimated what simulation, compute power, and static datasets could accomplish. The core issue is that autonomous vehicles fundamentally change how other drivers, cyclists, and pedestrians behave around them, something no simulation can fully predict.

The hardest problems in autonomous driving involve everyday ambiguity that doesn't show up in controlled testing scenarios. A scooter cutting across a pickup point, a driver edging into a gap, or a pedestrian's hesitation at a crosswalk can test a robotaxi far more than a clean, ideal-road scenario. These real-world interactions reveal gaps that world models, advanced AI systems designed to understand cause and effect, cannot fully replicate.

"Launching a robotaxi is becoming easier but building one that can scale commercially and operate safely remains difficult," stated Dr. Tiancheng Lou, founder and Chief Technology Officer at Pony AI.

Dr. Tiancheng Lou, Founder and Chief Technology Officer at Pony AI

What Does Real-World Road Exposure Actually Teach Robotaxis?

Lou argues that companies must move beyond supervised driver-assistance systems and operate fully driverless vehicles across complex, scaled urban environments to truly compete in the Level 4 autonomous market. Level 4 refers to vehicles that can drive themselves in most conditions without human intervention. Commercialization depends as much on fleet operations as on the underlying technology, because companies need enough real-world interactions to continuously improve their systems.

World models can help systems understand how other road users react when a robotaxi slows down or behaves cautiously at a junction, but those models cannot replace actual road exposure. The distinction matters because it suggests that companies with the largest, most diverse fleets operating in the most challenging environments will have a structural advantage over competitors relying primarily on simulation.

How to Evaluate Robotaxi Progress Beyond Lab Metrics

• Real-World Fleet Size: Companies operating larger fleets across more cities and diverse traffic conditions accumulate more data and edge cases, providing a competitive advantage that simulations cannot match.
• Operational Complexity: Robotaxis tested in dense urban environments with heavy pedestrian traffic, informal signals, and unpredictable behavior face harder challenges than those in controlled or suburban settings.
• Safety Track Record: Actual incident data and passenger feedback from commercial operations reveal system weaknesses that simulations miss, enabling faster iteration and improvement.

The robotaxi race is heating up globally. Waymo, backed by Alphabet, remains the U.S. market leader with its Waymo One service providing more than 250,000 paid trips weekly across Phoenix, San Francisco, Los Angeles, and Austin, with plans to expand to Atlanta, Miami, and Washington. Tesla is also pushing into robotaxis, with CEO Elon Musk touting a future large autonomous fleet, though some investors have criticized Tesla's Full Self-Driving system as problematic compared to Waymo's performance.

Pony AI operates its own robotaxi service called PonyPilot in major Chinese cities and has also developed Level 4 autonomous trucking technology. The company's emphasis on real-world road exposure aligns with a broader industry shift away from pure simulation-based development toward hybrid approaches that combine lab work with extensive field testing.

Lou's argument carries weight because it challenges a common assumption in AI development: that more compute power and better algorithms can solve problems in isolation. In robotaxi development, the messiness of real traffic, the unpredictability of human behavior, and the infinite edge cases that arise on actual roads cannot be fully captured in any simulation. Companies that recognize this and prioritize real-world fleet operations over lab benchmarks may find themselves better positioned to scale safely and profitably in the coming years.