The 2nd Generation VLA, through model scaling and architecture upgrades, begins to connect the past, present, and future: remembering longer road information, understanding environmental changes in real time, and predicting what might happen next.

The same set of Physical AI capabilities also begins to enter vehicle intelligence, cover more vehicle models, and extend to more embodiments such as robots.

01

Technological Breakthrough: AI Begins to Connect the Past, Present, and Future

From model breakthroughs to systemic evolution, Physical AI enters the real world. What Physical AI ultimately needs to solve is: understanding the world and acting in it. Model paradigms set the direction, model scale raises the capability ceiling; data provides experience for understanding the real world; cloud and on-device computing power support training and real-time operation respectively; embodiments allow AI to enter the real world for continuous validation.

The biggest change this time is that the model begins to understand "time." A single image can only tell the model "there is a person here," but cannot fully explain where he came from and where he is going. The 2nd Generation VLA no longer only looks at the current frame, but lets continuously occurring information participate in judgment together. Seeing the world takes only an instant, but understanding the world takes time.

Connect the Past, Understand the Present, Predict the Future

Infini-VLA: Let past effective information continuously participate in current judgment
· Streaming Inference: continuously receive new information and update decisions in real time
· X-Foresight + Flow Matching: deduce multiple future possibilities and generate more natural driving trajectories
· MoT Hybrid Architecture: allow different scenarios to invoke different experts while retaining cross-scenario visual information

These capabilities together help the model understand a continuously changing physical world.

02

Product Implementation: From Model Capabilities to Complete Tasks

Larger model, longer memory, faster response, farther foresight
On-device model parameters increased by 3.5 times, supporting up to 30 seconds of effective temporal sequence, and can predict multiple possible changes within the next 6 seconds. The vehicle can combine traffic light changes, movement trends of traffic participants, and road environment to make judgments earlier and respond faster. What you ultimately feel is: when facing occlusion, sudden crossing, cut-ins, and complex intersections, the vehicle acts earlier, decisions are more coherent, and the ride is smoother.

VLA+VLM Integration, Master Agent Onboard
The Master Agent based on the self-developed Omni multimodal foundation can understand complex intentions, decompose tasks, and coordinate vertical agents such as intelligent driving, chassis, cockpit, and body control for collaborative execution. When the user says "find a place to stop," the vehicle not only recognizes a voice command, but understands the need, judges the environment, plans the task, and completes parking, achieving natural language from intention directly to Action. True vehicle intelligence is not just understanding commands, but getting things done completely.

One foundation enables capabilities to break through upward, cover downward, and extend outward
Upward, the same-origin technology validates higher-level autonomous driving capabilities in Robotaxi. Currently, it has completed over 2,000 internal test rides and obtained the qualification for driverless road testing (no safety driver). Downward, through same-origin distillation, VLA 2.0 Lite enters platforms with lower computing power, covering more vehicle models and users. The stronger the foundation, the higher the capability ceiling, and the higher the starting point of distilled capabilities.

03

Capability Extension: From Automobiles to a Broader Physical World

A single model version determines one experience, but an Infra determines long-term evolution
The upgrade of the 2nd Generation VLA is not an isolated breakthrough; behind it is an AI Infra covering data, training, simulation, evaluation, deployment, and optimization. Real fleets continuously provide new problems, the data engine discovers high-value and long-tail samples, the training system drives model iteration, and the simulation platform validates model behavior at scale. Single training data throughput has increased 10 times compared to half a year ago, allowing more real-world problems to enter the training and validation loop faster. The real moat is not this version of the model, but the ability to continuously train the next version.

From Automobiles to Robots, the Physical AI Flywheel Begins to Compound
Improved model capabilities bring better product experience; more mass-produced products enter the real world, generating new data and problems; data undergoes training, simulation, and validation, then drives the evolution of the next generation of models. As this closed loop continues to operate, capabilities will no longer be limited to a single version, a single vehicle model, or a single embodiment, but will continuously extend to automobiles, Robotaxis, robots, and more intelligent products. What truly compounds is not a particular version of the model, but making every embodiment the starting point for the next round of evolution.

As a premier event in the automotive technology and engineering field, AMTS is committed to building a cross-border empowerment platform for high-end intelligent automated manufacturing driven by "new quality productive forces," focusing on the entire chain of R&D, design, and manufacturing, offering a one-stop gathering of global advanced manufacturing and assembly testing solutions. Throughout the year, it builds a one-stop exchange platform for system integrators, Tier 1 and Tier 2 suppliers, and vehicle and parts R&D and manufacturing engineers in the automotive technology and engineering field worldwide.

Looking forward to meeting you again around the world——

Guangzhou · Oct 23-25 —— 2026 World Automotive Productivity Conference

Bangkok · Nov 18-21 —— Thailand Automotive Engineering & Automation Industry Tour

Serbia · Nov 17-19 —— Central and Eastern Europe Ecosystem Partner Day

Shanghai · July 7-9, 2027 ——  AMTS 2027 | The 22nd Shanghai International Automotive Manufacturing Technology & Material Show