XPeng VLA 2.0: New AI model adds 30-second memory and six-second prediction

XPeng has announced a major upgrade to its second-generation VLA (Vision-Language-Action) model, with the updated system due to roll out to Ultra and Ultra SE models in September. The G9L will be the first model to receive the new version.

The distilled VLA 2.0 Lite is also scheduled to begin rolling out in September, with the G9L Max as the launch model.

The biggest change in the second-generation VLA is the addition of a time dimension to XPeng's physical-world model. Instead of working only with what is happening in the current frame, the system can now use information from earlier moments to understand how a scene is changing over time.

The new Infini-VLA long-sequence architecture can retain up to 30 seconds of scene information, giving the assisted-driving system a longer context window when making decisions.

XPeng's X-Foresight predictive world model has also been added. It can look up to six seconds ahead, modelling how surrounding road users are likely to move.

The different systems handle different parts of the process: Infini-VLA retains recent scene information, streaming autoregressive reasoning processes the current situation, while X-Foresight and Flow Matching are used to predict what happens next.

XPeng says the updated second-generation VLA improves its multi-dimensional safety capability by 20 times. The edge model has also grown by 3.5 times, with more than 15 times the parameters of mainstream VLA models, according to the company.

The upgrade also extends beyond assisted driving. XPeng has introduced Master Agent, bringing its VLA and VLM systems together across the vehicle. The company says this allows the car to interpret what a user is trying to achieve rather than simply responding to individual voice commands.

The same technology stack is designed to scale up to L4 automated driving, while model distillation allows it to run on hardware with lower computing requirements and be deployed across a wider range of production vehicles. XPeng also says the architecture is designed to generalise across different overseas driving environments.

Through Token compression and distillation training, XPeng is reducing the number of tokens needed for visual processing while retaining the information used by the larger model. The resulting VLA 2.0 Lite can therefore run on less powerful hardware, allowing more production vehicles to use the system.

The same AI platform is also being used in XPeng's robotics programme. Its general-purpose humanoid robot IRON uses three Turing AI chips with a combined computing power of 2,250 TOPS. XPeng says its physical-world model can run directly on the robot, allowing it to handle complex tasks without remote control while keeping latency low and data processed locally.

XPeng will continue to expand the system through its upcoming XOS 6.3.0 update. The edge model will again see a 3.5-times increase in parameters, while the longer memory window is intended to give the AI driver more information when dealing with changing traffic situations.

The new Master Agent will also bring VLA and VLM together across the vehicle, covering both driving and cabin functions. XPeng says some of the L4 capabilities developed for its Robotaxi programme will also be brought to production cars.

Web: http://www.bitauto.hk/

YouTube: http://www.youtube.com/@bitautohk

Facebook: https://www.facebook.com/bitautohongkong/