Tsinghua's Li Shengbo Analyzes Three Major Challenges of End-to-End Autonomous Driving

[Technical Insight] Professor Li Shengbo of Tsinghua University pointed out that China's end-to-end autonomous driving development faces three core bottlenecks.

Key Trends: Data, Computing Power, and Algorithms Are Primary Constraints

Li Shengbo emphasized at the High-Level Forum on Intelligent and Electric Vehicle Development on April 12, 2026, that current end-to-end autonomous driving systems are constrained by insufficient data scale, high computing costs, and limited algorithm maturity.

Strategic Foundation: Simulation Plus Efficient Algorithms Are Crucial to Breakthroughs

Since 2018, Tsinghua University has been advancing a two-stage end-to-end model, integrating high-fidelity simulation software with real-vehicle data. In 2024, the team completed China’s first open-road test based entirely on a neural network architecture. The team advocates leveraging simulation to scale up data generation and reduce reliance on raw data collection and computing infrastructure expansion.

Industry Impact: Embodied Intelligence Complexity Is Severely Underestimated

Compared to autonomous driving, training embodied intelligence robots is 5–10 times more difficult, requiring tens of billions of data segments and approximately 100 billion parameters. The industry broadly underestimates this complexity. Li Shengbo predicts that the field of physical intelligence will drive the emergence of numerous new technologies and companies over the next 10–15 years.