Xpeng's new VLA model can "see" 6 seconds into the future
During their recent Physical AI sharing session, Xpeng demonstrated how this temporal awareness moves beyond simple object detection into actual behavioral understanding.
Breaking the "Single Frame" Limitation
Traditional autonomous driving systems often struggle because they react to what is happening right now. If a car in front of you suddenly swerves, a system without temporal memory might only see the new position, whereas a human driver remembers the trajectory that led to that swerve.
Xpeng is tackling this through three specific technical pillars:
- Infini-VLA (The Memory): This component handles long-sequence temporal data. While edge computing constraints usually limit how much a car can "remember," Xpeng has optimized this to a 30-second window. In a demo, a vehicle encountered a car making a U-turn. Instead of seeing a momentary gap in traffic and accelerating into it, the model recognized the ongoing process of the U-turn and waited for the maneuver to complete.
- Streaming Inference (The Reflex): To prevent the lag that comes with processing heavy temporal data, they implemented streaming autoregressive inference. This allows the model to output trajectories while simultaneously receiving new inputs. Xpeng claims this has boosted decision-making speeds by 300%, mimicking the way a human eye scans the road while the brain simultaneously prepares muscle movements.
- X-Foresight & Flow Matching (The Intuition): This is the predictive layer. By using historical states to simulate potential future trajectories, the model performs a type of "defensive driving." It doesn't just see a car in the oncoming lane; it predicts whether that car is likely to yield, continue forward, or cut in, allowing the vehicle to adjust its speed before a conflict even occurs.
Scaling Up for Real-World Chaos
One of the most impressive (and risky) parts of this update is the massive scale-up in model parameters. Xpeng has increased the model size by 3.5x to handle "corner cases" that fixed-rule systems fail at.
They shared footage of a vehicle navigating a ferry terminal—an environment with no painted lane lines—and a narrow mountain road with a low-hanging tree branch. Unlike some competitors that might see an obstacle and simply slam on the brakes, the VLA model actually "reasoned" through the space, slowing down to calculate if the chassis could clear the branch.
To feed this massive model, Xpeng is using a massive data loop:
- Data Throughput: Current training involves 100 million video snippets.
- Vectorized Retrieval: When a new edge case is found, the system searches a database for semantically similar scenarios to build a targeted "error collection."
- Massive Simulation: Their simulation testing volume increased by 290% recently, varying weather, lighting, and angles to stress-test the model before it hits the road.
The Shift from "Car Company" to "Physical AI"
The technical roadmap is clear: Xpeng is moving toward a "Car as Robot" philosophy. They are unifying the architecture so the same underlying foundation can power everything from L2 driver assistance to L4 Robotaxis and even humanoid robots.
The new VLA 6.3.0 version will begin rolling out in September, starting with the Ultra and Ultra SE models. For those on the single Turing chip Max models, a distilled "VLA Lite" version is also coming.
While the industry is currently locked in a brutal price war, the real battle is shifting toward who can build the most capable "Physical AI." Xpeng is betting that by giving cars a sense of time—the ability to learn from the past and anticipate the future—they can move beyond mere automation and into true intelligent agency.