QCraft CEO says 'physical AI' is key to closing self-driving gap

At a time when artificial intelligence can defeat world champions in board games yet still struggles to match human drivers on open roads, Yu Qian, co-founder and CEO of autonomous driving firm QCraft, argues the gap reflects a deeper technological shift now underway.

Speaking on Thursday at an industry forum in Munich co-hosted by the Technical University of Munich and the Center Automotive Research, Yu framed the next phase of competition as "physical AI" — systems capable of reasoning in the real world rather than optimizing within closed environments.

A decade ago, AI systems such as those used in Go benefited from near-infinite, low-cost trial-and-error in virtual settings. Autonomous driving, by contrast, operates in an unpredictable physical environment where safety constraints sharply limit experimentation.

"That's the fundamental bottleneck," Yu said in a discussion with Ferdinand Dudenhoeffer, founder of the Center Automotive Research. "You cannot let systems 'learn by failure' on real roads."

Yu outlined a three-stage evolution of AI, moving from imitation to human-like intelligence and now toward what he described as "superhuman" capabilities.

The transition hinges on whether machines can internalize physical laws, social norms and causal reasoning — not merely replicate observed behavior.

QCraft's approach centers on combining world models with reinforcement learning.

The company likens the system to a virtual driving school, where simulated environments generate millions of long-tail scenarios — from complex urban intersections to extreme weather — while algorithms iteratively refine decision-making without real-world risk.

That architecture underpins what Yu calls a shift from "passive memory" to "active reasoning," allowing systems to generalize beyond training data.

The strategy has been tested at scale: as of January, the company's assisted-driving systems had been deployed in more than 1 million vehicles, providing continuous real-world feedback loops.

Unlike some rivals pursuing higher computing power, QCraft emphasizes efficiency. Its urban navigation-on-autopilot functions run on a 128 TOPS platform, handling tasks such as unprotected turns and dense nighttime traffic.

Yu said the approach supports a "mass-market autonomy" model rather than a premium-only rollout.

Commercially, the company is pursuing a dual-track strategy, pairing advanced driver assistance with fully autonomous applications.

Driverless logistics vehicles are already in operation, while robotaxi pilots are scheduled to begin in 2026 ahead of broader deployment.

Munich has become a focal point for its global expansion, with an office established in 2025.

Yu said the company aims to integrate experience from China's complex traffic environments with Germany's engineering expertise.

Compatibility with major chip platforms and compliance with international safety standards are intended to ease adoption by global automakers.