Research on Automotive and Robot VLA: Hybrid Architectures Become Mainstream, VLA Integrates with General World Models, and Reinforcement Learning Serves as Core Engine

Vision?Language?Action (VLA) model is a model integrating vision, language and action modalities. Adopting a unified multimodal learning framework, it integrates perception, reasoning and control, and generates executable physical world actions (e.g., robot joint motion, and vehicle steering/acceleration/braking control) directly from visual inputs (images/videos) and language instructions.   

VLA equals VLM (for understanding and description) plus E2E (end?to?end decision and control) plus CoT (chain?of?thought human?like reasoning). In terms of capability comparison, VLA delivers precise 3D perception, commonsense understanding, logical thinking and interpretability simultaneously.

1. Status Quo and Major Challenges of VLA
The evolution of VLA in autonomous driving falls into four stages.
Language as interpreter (Pre?VLA): Language models only generate scene descriptions without participating in control.
Modular VLA: Language acts as a planning component for decision, yet multi?stage workflows incur latency.
Unified end?to?end VLA: Sensor inputs are directly mapped to actions via a single forward propagation.
Reasoning?enhanced VLA: LLMs enter the control closed loop to enable long?term reasoning, memory and interaction capabilities, exemplified by Li Auto MindVLA.

Implementation timeline:
Segmented end?to?end came into mass production from 2024 to 2025.
One-model end?to?end and VLA were largely rolled out between 2025 and 2026.
The year 2026 marks a critical window period of “intensified multi?route competition and accelerated paradigm integration” for intelligent driving large models.

Technical status and core indicators.

Wide?ranging model parameters: NVIDIA Alpamayo 1.5 includes 0.5B/10B parameters; DeepRoute.ai uses a 40B?parameter foundation model; StepVL foundation model from Afari Technology features 32B parameters (distilled to 7B and 3.6B); Li Auto MindVLA 32B?parameter foundation model is distilled into a 3.6B?parameter MoE variant, 4B for vehicle model.     

In?vehicle real?time performance: Li Auto leverages sparse attention + MoE to realize 10Hz and 100ms latency on Orin X; Xpeng’s second?generation VLA enables <80ms latency; DeepRoute.ai’s 40B?parameter model uses KV Cache, Multi?Token Prediction (MTP), quantization and customized engine to achieve single?step latency of 60?85ms and 10?15Hz closed loop. 

Computing power adaptation: DeepRoute.ai can deploy pure driving VA models on 100TOPS platforms and reasoning?capable VLA models on 500TOPS platforms; Leapmotor D19 equipped with dual Qualcomm 8797 chips (1280TOPS) realizes end?side VLA?assisted driving; Geely H9 adopts dual NVIDIA Thor chips (2000TOPS).

Open?loop performance: Based on the nuScenes dataset, VLA exhibits notably lower trajectory errors than world models. For instance, AutoDrive?R2 with 7B parameters delivers an L2 distance of 0.19m and SENNA achieves 0.22m; the best?performing world model Drive?OccWorld reaches 0.32m. Such results demonstrate VLA’s ability to reproduce human real?world driving trajectories.   

Major challenges

Real?time performance and computing power bottlenecks: Traditional auto?regressive VLA generation only reaches 3?6Hz, and single?reasoning latency commonly exceeds 200ms, consuming a lot of vehicle computing power and storage bandwidth.

Data supervision deficit: VLA receives high?dimensional visual inputs yet is supervised by low?dimensional sparse actions, limiting model potential. World models are required to predict future images for dense supervision, or video prediction pre?training shall be adopted.

Lack of safety redundancy: End?to?end single?model VLA has low fault tolerance and requires fallback from traditional algorithms, e.g., the widely adopted fast?slow dual?system, Horizon Robotics Lite Safety Checker and Bosch safety gating reward mechanism.

Hallucinations and long?tail scenarios: Success rates drop in complex or unseen scenarios. World model + reinforcement learning exploration is required to improve, e.g., Bosch ExploreVLA, Huawei WEWA and Momenta R7. 

2. Typical VLA Solutions of OEMs
Among OEMs, Xpeng’s second?generation VLA and Li Auto MindVLA stand as representative cases, corresponding respectively to two technical routes, namely “native multimodal physical world foundation model” and “space?language?action unification + implicit world model”. 

Xpeng’s Second?Generation VLA
Xpeng holds to the view that intelligent driving is essentially a physical AI problem. Its second?generation VLA is built as a native multimodal physical world foundation model. A native multimodal tokenizer enables highly efficient early?stage fusion to avoid single?modality bias. Visual reasoning chain?of?thought (CoT) boosts reasoning efficiency by 32 times. In car following scenarios, the model automatically generates maneuver proposals such as lane change or car following, and produces abstract bird?eye?view diagrams for scoring. Native cockpit?driving linkage allows the model to generate not only actions but also videos and sounds, serving as the foundation for VLA and the foundation framework for world models, simulation, and reinforcement learning. Referring to two?stage models, it retrains one-stage foundation models and eliminates language translation links to bring latency below 80ms.        

Li Auto MindVLA
Its architecture consists of three components.
V (spatial intelligence): Based on BEV and OCC, it adopts 3D Gaussian as intermediate representation, leverages LiDAR point clouds as 3D geometric prompts, and performs 3D scene reconstruction via 3D ViT encoder and feed forward 3DGS. Static environments and dynamic objects are modeled separately.
L (language intelligence): Retrains LLM foundation model (leveraging MoE and sparse attention mechanism). Fast?thinking parallel decoding directly outputs Action Tokens, while slow thinking outputs CoT and Action Tokens simultaneously. 
A (action strategy): Adopts VLA?MoE architecture embedded with Action Expert. Discrete diffusion and parallel decoding iterative optimization are used to output high?precision driving trajectories.

Four?phase engineering:
VL foundation model pre-training (32B, post-distilled to 3.6B MoE to adapt to dual Orin-X/Thor-U)
Imitation learning post-training (4B)
Reinforcement training (RLHF + pure RL, building a closed-loop world simulator)
Driver agent HMI

Subsequent MindVLA?U1 enables unified streaming and joint modeling of language and continuous actions, and introduces Intent?CFG. 

Other OEMs:
Xiaomi XLA Cognitive Large Model (VLM + edge-cloud integration, VLA planned);
Leapmotor LEAP 4.0 VLA (edge-side full-modality, "understanding-planning-preview-judgment-correction" closed loop);
Great Wall CP Master Dual VLA (left brain driving agent, right brain cockpit agent);
Chery Falcon 900 (VLA + world model, supporting L3).

3. Typical VLA Solutions of Suppliers

The most typical solutions of suppliers are NVIDIA Alpamayo and DeepRoute.ai 40B VLA, representing “reasoning?based dual?LLM VLA + world model” and “40B unified base three-stage VLA” solution respectively.  

NVIDIA Alpamayo
NVIDIA Alpamayo is a system?level VLA solution encompassing physical AI dataset, VLA large model and AlpaSim simulation framework. It provides OEMs with three generations of trajectory generation options for vehicle deployment: regression prediction (TensorRT, e.g., SparseDrive), diffusion generation (TensorRT, e.g., DiffusionDrive), and flow matching (TensorRT?Edge?LLM, e.g., Alpamayo based on Qwen3 VL).

Take Alpamayo?R1?10B as an example. Inputs include historical images, user instructions, historical trajectories and noisy actions. Inputs are converted into Text, Image and Trajectory Tokens via VLM Flow Matching Tokenizer. The 8B?parameter Qwen3 VL?LLM handles scene understanding and implicit CoT reasoning, outputs reasoning texts and generates KV Cache. The 2B?parameter Qwen3 VL?LLM receives noisy trajectories and KV Cache, conducts progressive denoising and correction via Flow Matching, and finally outputs future trajectories. Cloud Cosmos world model generates training data for long?tail scenarios. Vehicle deployment adopts traditional algorithm as safety fallback + end?to?end VLA as primary system, supporting L2?L4. Future latent space reasoning is expected to accelerate by 2?4 times.    

DeepRoute.ai 40B VLA
DeepRoute.ai breaks down autonomous driving decision into three phases.
Observation phase - Multi?camera videos are encoded into approximately 1,000 visual tokens.
Reasoning phase - The model conducts in?depth semantic analysis of scenarios and generates descriptions of key events and decision logics, with the number of reasoning tokens strictly controlled within 10?50.
Execution phase - Outputting driving control commands requires only about 10 tokens.

It integrates three capabilities of "driver (acting based on sensor inputs), analyst (analyzing causality), and commentator (judging and making decisions)" through a unified 40B-parameter foundation model. Joint training is implemented across three task categories, namely, V+A, V+A→L and V→L+A, to realize “thinking before driving”. Pre?training switches from trajectory supervision to video prediction. Massive videos are leveraged to learn physical laws at per?pixel level, lifting data utilization rate from 0.001% to 100%. During deployment, KV Cache, MTP, quantization and customized reasoning engines reduce single?step latency to 60?85ms to achieve a 10?15Hz vehicle real?time closed loop. Model distillation is performed according to computing power: pure driving VA models run on the 100TOPS platform, and complete VLA models operate on the 500TOPS platform.        

Other suppliers:
?Afari Technology adopts the ““VLA+E2E” collaborative closed loop. The VLA slow system outputs CoT texts while the E2E fast system outputs target detection/lane detection results for fusion into planning and control. StepVL 4.0 is distilled from 32B pre-trained to 7B.  
?QCraft upgrades to the "VLA + world model + reinforcement learning" unified architecture in 2026, and realizes urban NOA on single Journey 6M chip.
?Zhuoyu launches VLA World Model (native multimodal foundation model, Chain of World, and structure-motion decoupled latent motion representation).

4. Development Trends of VLA

Trend 1: Hybrid Architectures Become Mainstream

Deep integration of Diffusion, Transformer and LLM/VLM becomes mainstream. Diffusion excels at generating high?quality continuous actions and trajectories. Transformer is good at long sequence modeling. LLM/VLM is skilled in semantic and multimodal understanding. Representative examples include Li Auto MindVLA (3D Gaussian + MoE LLM + Diffusion Action Expert), NVIDIA GR00T-N1 (Fast-Slow Dual System: Fast 200Hz Diffusion Action, Slow 10Hz VLM), HybridVLA (Autoregression + Collaborative Diffusion). The industry has formed three integration models: ① One-model end-to-end + world model + RL (Momenta, Horizon Robotics); ② VLA + world model (XPeng, etc.); ③ E2E + VLM/VLA foundation model (Afari Technology VLA slow system + E2E fast system).   

Trend 2: VLA and General World Model Integrate into World VLA / VLA World Model

VLA undertakes cognition and action while the world model takes on future prediction. Their unification transforms automobiles from transportation means into mobile robots, shifting from rule?driven to cognition?driven, with capabilities of autonomous perception, reasoning & decision and precise execution. Zhuoyu’s VLA World Model has evolved into its third-generation native multimodal foundation model. With Chain of World, it performs multistep world state prediction in latent space, achieving "thinking before acting." Structure-motion decoupling and latent motion representation reduce reconstruction costs. Geely G-ASD integrates VLA and world model, enabling vehicles to automatically perform tasks. WorldVLA jointly understands actions and images for generation, with the world model and action model mutually reinforcing each other.        

Trend 3: VLA + World Model + Reinforcement Learning (RL) Trinity Integration, with RL as the Core Engine   
 
The industry forms a "pre-training → simulation → reinforcement learning" three?layer architecture. The world model generates long?tail scenarios, VLA conducts in-loop reasoning, and reinforcement learning iterates optimal strategies in the inference space. Representative examples include: Huawei WEWA 2.0 (multi-agent gaming + cloud online RL, training intensity increased by 10 times); Momenta R7 (three-stage process: pre-training → simulation → RL, turning AI from "imitator" to "decision-maker"); Pony.ai’s PonyWorld 2.0 (self-diagnosis + targeted evolution + precision flywheel).  

Meanwhile, world models evolve from pixel?level prediction toward latent space and causal reasoning. NVIDIA Alpamayo achieves 2?4?fold acceleration via implicit reasoning in the latent space, and generates a complete reasoning chain through Chain of Causality (CoC). Li Auto embeds predictive implicit world models into VLA. Xpeng eliminates language translation links and revises architecture from V?L?A to V/L?A to mitigate information loss. Huawei DriveVLA?W0 verifies that with world model integration, collision rates keep decreasing as data volume expands from 0.7 million to 70 million frames and such advantages are amplified, strengthening the data scaling law. 

Trend 4: Engineering Implementation and Safety Assurance Accelerate

One-model end?to?end and VLA solutions are largely implemented from 2025 to 2026. The evolution of L3/L4 has driven safety redundancy to become a necessity (traditional algorithm fallback + end-to-end main system, e.g., NVIDIA's fast-slow dual systems, Horizon Robotics' Lite Safety Checker, and Bosch's safety gating PDMS reward). Hierarchical distillation of computing power has become key to mass production: flexible deployment of VA/VLA (DeepRoute.ai) at 100-500 TOPS, and high-performance computing platforms such as dual Thor/dual 8797 supporting L3.    

VLA serves as the core route for intelligent driving to evolve from “end?to?end perception?control” toward “understanding?reasoning?control”. In 2026, driven by both OEMs (XPeng, Li Auto, etc.) and suppliers (NVIDIA, DeepRoute.ai, Afari Technology, QCraft, etc.), VLA is deeply integrated with world model and reinforcement learning, forming a hybrid architecture, and the prototype of World VLA takes shape. As latency, computing power, data supervision, and security redundancy issues are gradually resolved, VLA will support the large-scale deployment of L3 and above autonomous driving and enable vehicles to evolve into general agents in the physical world.