XPeng Unveils Second-Gen VLA with 30-Second Memory and 6-Second Prediction, Rollout in September

At the sharing session, XPeng gave its in-car model what it calls an "Eye of Agamotto": looking backward, it captures 30 seconds of history; looking forward, it projects what is likely to occur in the next six seconds. The model uses accumulated information from the past moments to calculate several higher-probability future states before choosing an action for the current instant. For assisted driving, having a timeline of memory enables more human-like behavior, since human drivers never only look at the present frame.

The architecture consists of multiple components. Infini-VLA handles memory, with the production version set to retain 30 seconds of road history. In a demo video, when a vehicle ahead made a U-turn, the test car did not immediately accelerate even as a gap appeared; it waited for the other car to complete the maneuver, understanding the ongoing action rather than just a single frame of open space. Streaming Inference, or "streaming autoregressive inference," lets the model input, reason, and output trajectories simultaneously instead of waiting for the full history, improving decision speed by 300% according to XPeng. X-Foresight extrapolates possible states over the next six seconds, and Flow Matching fits a distribution of possible trajectories to help select the appropriate action.

The new VLA expands model parameters by 3.5 times. In another demo, the vehicle drove through a ferry terminal without clear lane markings, found the entrance, boarded the ferry, and exited after docking. On a narrow mountain road, it encountered a hanging branch and slowed down to assess the gap between the branch and the car rather than stopping or blindly proceeding. XPeng contrasted this with a recent case where a Tesla running FSD v14.3.7 drove from the second floor to the sixth floor of an apartment parking structure and kept approaching a metal wire fence at the top until the owner took over. The company argues that a model must understand whether a space is a drivable path, not just perceive space.

To support the larger model, XPeng says its single training run now ingests 100 million video clips, and a vectorized data retrieval system lets engineers pull semantically similar scenarios to build targeted "error sets." Simulation varies vehicle models, weather, lighting, and viewpoints to repeatedly validate the same model. Daily simulation-validated new models increased 290% from June to the two weeks prior to the event, and overall safety capability in simulation and testing improved more than 20-fold, according to the company.

For vehicles with a single Turing chip, XPeng used Hybrid ViT to redesign the most computationally intensive visual components, preserving as much of the VLA's underlying capability as possible. A distilled Lite version, second-generation VLA Lite, will begin rolling out in September, with the XPeng G9L Max as the first to receive it.

XPeng also rebuilt the vehicle's on-board brain. Liu Xianming, head of XPeng's General Intelligence Center, defined the approach as "Car as Robot." The Omni all-modal model combines language, environmental, and contextual understanding, while Master Agent breaks intents into executable tasks. Unlike a system that handles single orders, the new architecture tracks the entire dialogue and understands the shared goal. For example, when a user says "find a place to pull over," the system must change lanes, observe barriers and bus stops, eliminate unsuitable spots, and then stop in an appropriate location.

The same architecture and model are intended for both L2 driver assistance and L4 Robotaxi. XPeng disclosed that its Robotaxi internal testing has completed more than 2,000 rides, and a Robotaxi equipped with the second-generation VLA recently obtained a remote-testing permit from Guangzhou, allowing road tests without a safety officer in the driver's seat. A video shown at the event displayed a robot performing indoor navigation and manipulation, which XPeng said uses the same base model, consistent with its self-definition as a physical AI company.

XPeng is not alone in linking cars to embodied intelligence. Tesla has placed cars, Robotaxi, Optimus robots, and AI training in a single blueprint, and Li Auto has shifted its positioning to an embodied-intelligence enterprise. For automakers, building a car increasingly means building a robot that acts in the physical world. While the pivot to "physical AI" also offers a more compelling capital story, XPeng's path requires sustained investment, repeated validation, and mass-produced versions. Whether the new identity holds will depend on whether car owners from September onward feel that the vehicle remembers more, judges earlier, and reacts faster.