XPENG unveils X-Mind autonomous driving brain at CVPR 2026 Workshop

XPENG has introduced X-Mind, a new Predictive World Model framework designed to enhance the reasoning and forecasting capabilities of autonomous driving systems. Unveiled at the CVPR 2026 Workshop, the technology utilizes a Visual Chain-of-Thought approach to simulate future traffic scenarios before a vehicle makes a decision. This development marks a significant step in XPENG's Physical AI roadmap, aiming to transition from reactive perception-to-action systems to proactive, human-like driving intelligence.

Xianming Liu, Head of XPENG Group's General Intelligence Center, presented the X-Mind framework as the latest evolution in the company's autonomous driving roadmap, following previous iterations like X-World and X-Foresight. X-Mind is characterized as a Predictive World Model that enables vehicles to understand how the environment evolves in response to their actions. By integrating proactive reasoning and long-horizon forecasting, the system moves beyond traditional reactive models, allowing autonomous vehicles to anticipate traffic changes through internal simulation rather than simply responding to current sensor data.

The technical architecture of X-Mind relies on three core innovations: Thought Sketch, Recurrent Block Diffusion (RBD), and Visual Chain-of-Thought (Visual CoT) visualization. Thought Sketch creates cognitive representations by combining Bird's-Eye-View (BEV) layouts with driving priors to preserve essential road structures and navigation intentions while minimizing computational load. Meanwhile, RBD addresses the latency issues typically associated with diffusion methods, allowing for high-quality scene generation within a single forward pass. This ensures that the model's advanced reasoning capabilities can be deployed in real-time on standard automotive-grade chips.

Trained on hundreds of millions of real-world driving data frames, X-Mind has demonstrated significant improvements in trajectory prediction accuracy and the handling of complex long-tail scenarios. The inclusion of Visual CoT visualization provides a transparent look into the model's decision-making process, showing how it predicts obstacle movements and lane connectivity before executing a maneuver. This transparency is critical for system validation and safety. Together with X-World and X-Foresight, X-Mind completes XPENG's Physical AI foundational model roadmap, positioning the company to deliver more sophisticated, human-like performance in the competitive autonomous vehicle market.