The Era of Physical AI: Vehicles Will Evolve into Embodied Intelligent Agents on Wheels
NVIDIA divides the AI industry development into five advanced stages, clearly outlining the evolution path of artificial intelligence from digital perception to closed-loop control in the physical world. These stages are: single image classification AI (AlexNet), perceptual AI, generative AI, agentic AI, and Physical AI. AlexNet, born in 2012, laid the foundation for the industrialization of deep learning, but it could only complete fixed and single image classification tasks, belonging to typical weak AI. Perceptual AI achieved multi-modal information recognition such as voice, vision, and user preferences, but lacked the ability to generate, reason, and actively execute. Generative AI broke through the boundaries of content creation, capable of autonomously generating text, images, and videos, but could not interact with the real physical environment. Agentic AI can complete simple digital tasks, but still cannot exert a substantial impact on the physical world.
Until the advent of Physical AI, artificial intelligence stepped out of digital virtual scenarios. Through carriers such as autonomous driving vehicles and general-purpose robots, it directly perceives the real environment, makes autonomous decisions, and executes physical actions, completing the full closed loop from perception to action and achieving the deep integration of digital capabilities and physical entities. Among the numerous implementation carriers, the industry consensus is that the first half of Physical AI focuses on the large-scale implementation of autonomous driving, while the second half will usher in the comprehensive popularization of in-vehicle embodied intelligence.
Five Stages of AI Development: Autonomous Driving is Entering the Physical AI Stage
Source: Zosauto
The ultimate evolutionary direction of the automobile is to become a super intelligent agent on wheels, namely an in-vehicle AI robot: no longer just a means of transportation that takes people from Point A to Point B, but a mobile intelligent partner with omni-domain perception, short- and long-term sequential memory, and autonomous reasoning and understanding capabilities, capable of proactively perceiving needs and executing complex tasks in a closed loop. Meanwhile, as a crucial AI-native terminal, the vehicle will build a cross-scenario super entry point, connecting travel, home, and office scenarios, supporting users in fulfilling diverse needs such as office work and household affairs management while driving.
Evolutionary Direction of Vehicles: Super Intelligent Agents on Wheels
Source: Zosauto
Three Core Implementation Paths for In-Vehicle Proactive Intelligent Services
The core value of high-level in-vehicle embodied intelligence is to break free from passive command responses and achieve continuous, proactive, and closed-loop scenario-based services. The implementation paths can be summarized into three major routes:
01. Creating "One-Sentence Full-Process" Integrated Services for Single High-Frequency Scenarios to Solve the Pain Point of Fragmented In-Vehicle Functions
Abandoning scattered and independent functions, it encapsulates continuous tasks around the complete travel chain. For example, in a long-distance self-driving scenario, a single user command can trigger the entire closed loop: route navigation planning, dining reservations along the way, destination hotel booking, and parking space reservation, while simultaneously generating content materials such as itinerary wallpapers and Vlog scripts, achieving single-command-driven full-chain execution.
02. Achieving the Leap from "User-Triggered Commands" to "System-Proactively Pushed Services"
This is divided into two major supporting capabilities: First, deeply learning user behavior preferences, continuously accumulating data on travel times, destination tendencies, and operation habits to form a personalized service model that supports user intent prediction. Second, strengthening omni-domain scenario perception by integrating vehicle status (battery level, energy consumption, component operating conditions), external environment information (traffic conditions, traffic control, road events), and the physiological status of drivers and passengers (fatigue level, heart rate, etc.), triggering scenario-based proactive services based on comprehensive information.
03. Building Cross-Domain Linkage Capabilities among Cockpit, Autonomous Driving, and Cloud Ecosystems to Create Core Differentiation Barriers
Breaking the silos of various functional modules in the vehicle, it opens up data and command channels for cockpit interaction, autonomous driving systems, and third-party service ecosystems. For example, when a user drives from Shanghai to Hangzhou, the system predicts dining nodes based on time, matches the user's dietary preferences to recommend target restaurants and push discounts; simultaneously links navigation and real-time parking lot data to pre-occupy vacant parking spaces in malls; after the user confirms the order, the autonomous driving system automatically updates the navigation destination to the target parking lot, completing a seamless and continuous experience of "ordering - finding a spot - parking - dining". However, such high-level embodied intelligence and proactive cross-domain services place extremely high demands on in-vehicle edge-side inference capabilities. The current mass-produced in-vehicle hardware systems have obvious computing power bottlenecks, making it difficult to support the implementation of these capabilities.
Core Industry Pain Points: In-Vehicle Edge-Side Computing Power Bottlenecks and Mismatched Software-Hardware Cycles
Currently, it is difficult to implement high-level in-vehicle proactive intelligence and cross-domain collaborative services at scale. The core constraints lie in the insufficient edge-side computing power of vehicles and the mismatched iteration of software and hardware. At present, mass-produced edge-side models in the industry are generally lightweight models of 0.2B–1.3B. This is not due to limited algorithm capabilities, but because in-vehicle hardware computing power and real-time inference conditions form rigid constraints. 70%–80% of existing models in the market are equipped with mainstream cockpit chips such as Qualcomm 8155 and 8295. Among them, the 8295P has only 30 TOPS of NPU computing power, which can only support the operation of 1B–3B small models. It can only meet lightweight interactions such as basic wake-up, simple commands, and short-turn dialogues, and cannot carry high-level AI Agent capabilities such as multi-modal understanding, long-term sequential memory, and multi-task chain planning.
From the perspective of the whole vehicle architecture, the automotive EE architecture has the characteristics of long cycles, high solidification, and difficult iteration. Not only is the computing power of existing models insufficient, but a large number of brand-new mass-production platforms generally choose basic cockpit SoCs to control BOM costs. The core load of native cockpit chips is concentrated on screen rendering, audio and video decoding, and basic vehicle control. Computing resources are saturated all year round, making it impossible to stably support edge-side large models with high parameters, high concurrency, and strong reasoning. At the same time, the mass production cycle of in-vehicle chips is as long as 2–3 years, while the iteration cycle of edge-side large models is only 3–4 months. The natural mismatch in the life cycles of software and hardware leads to the problem that new cars face delayed intelligent iteration upon launch.
Optimal Industry Solution: External AI BOX Becomes the Core Solution for the Transition of In-Vehicle Agent Upgrades
Against this industry background, the external edge-side AI BOX, which requires no reconstruction of the whole vehicle architecture and no replacement of the original vehicle domain controller, has become the optimal incremental solution for automakers to quickly implement 3B–7B high-level edge-side large models and achieve large-scale upgrades of in-vehicle embodied intelligence. Through an independent external computing unit, it isolates the inference load of large models and releases the hardware pressure of the original vehicle cockpit. With the lowest modification cost and the shortest iteration cycle, it achieves a leapfrog upgrade in the capabilities of in-vehicle agents.
Currently, players deploying in the intelligent cockpit AI BOX track are divided into six major camps:
Represented by SenseAuto, integrated software and hardware AI solution providers rely on self-developed edge-side large models and cockpit agent frameworks to launch standardized SageBox hardware, natively equipped with mature in-cabin multi-modal perception capabilities. Represented by ThunderSoft's AIBOX-N1, AIOS manufacturers provide a complete software stack at the operating system layer, agent scheduling, and multi-modal interaction based on third-party chip hardware. Represented by Desay SV, Tier 1 complete machine solution providers focus on automotive-grade hardware and in-vehicle OS output, mostly adopting external cooperation models for large model capabilities. Chip manufacturers such as NVIDIA and Intel provide differentiated computing power foundations to empower AI-BOX hardware development. Pure software large model manufacturers such as ModelBest, Volcano Engine, and iFLYTEK do not develop hardware themselves, but rely on third-party AI-BOXes to carry models and interaction capabilities. There is also the OEM self-developed route, with representative products such as NIO's N-Box, building scalable external computing units based on the whole vehicle EE architecture.
As a modular incremental computing power carrier during the transition period of the whole vehicle architecture, the AI BOX can make up for the computing power shortcomings of the traditional cockpit domain, support the cockpit agent in achieving proactive prediction, continuous task scheduling, and cross-domain collaboration, and is an important hardware support in the evolution of vehicles into mobile intelligent partners on wheels.
Intelligent Cockpit In-Vehicle AI BOX Attracts Multiple Players Including Tier 1s, Chip Manufacturers, Large Model Companies, and OEMs
Source: Zosauto
In-Depth Analysis of Product Solutions from Leading AI BOX Manufacturers
Key AI BOX Manufacturer 1: Analysis of SenseAuto SageBox
SenseAuto made a significant release of SageBox at the 2026 Beijing Auto Show, providing a full-stack mass-production-level innovative solution for the evolution of autonomous agents. SageBox is an independent external AI computing unit equipped with NVIDIA Orin-X. It is positioned as a computing power enhancement solution for high-level vehicle agents. It does not replace the original vehicle cockpit SoC, and the original chip retains audio and video, display, and basic vehicle control functions. SageBox carries Sage edge-side large model inference and New Member agent task planning. The vast majority of cockpit complex commands and multi-modal interactions can be completed locally within the box, achieving offline availability and eliminating continuous cloud token costs. The cloud is only used as an optional enhancement to deal with extreme open-ended tasks that are difficult to cover on the edge side.
SenseAuto SageBox
Image Source: SenseAuto
SenseAuto SageBox has overall built a three-layer full-stack technical architecture of "edge-side large model + intelligent operating system + native in-vehicle agent". From top to bottom, it forms a complete closed loop of in-vehicle embodied intelligence, undertaking the core capabilities of cognitive thinking, system scheduling, and interactive execution respectively, to achieve high-level proactive cockpit intelligent services.
First Layer | Bottom Cognitive Foundation: Sage Edge-Side Multi-Modal Model (Sage 32B MoE)
As the core thinking brain of whole-vehicle intelligence, the Sage 32B MoE edge-side model undertakes the function of the in-vehicle cognitive foundation, possessing strong reasoning, multi-modal fusion, and long-term sequential task planning capabilities. It can efficiently complete the parsing of complex user intents, the decomposition of long-chain scenario tasks, logical decision-making planning, and multi-modal content understanding, supporting the capability upgrade from passive command response to proactive demand prediction. At the same time, the model is natively compatible with mainstream agent frameworks such as OpenClaw and Hermes, which can connect cross-scenario ecosystems such as travel and home, providing a universal computing power and algorithm foundation for the large-scale implementation of multi-domain edge-side agents.
Second Layer | Middle Scheduling Hub: Sage OS (Sage AIOS)
Sage OS is the core hub and system foundation of the three-layer architecture, undertaking the responsibilities of omni-domain resource scheduling and module integration. On the one hand, it is responsible for in-vehicle hardware computing power management, model inference scheduling, and task priority allocation, ensuring low-latency and highly stable local inference of the Sage large model. On the other hand, it connects the data flow and command flow among the underlying hardware, edge-side models, and upper-layer agents, providing a stable, secure, and standardized operating environment for the New Member native agent, achieving efficient collaboration between the "thinking layer" and the "execution layer".
Third Layer | Upper Interactive Execution: New Member 2.0 Native In-Vehicle Agent
As the upper-layer application, the New Member native agent is mainly responsible for executing interactions. It includes Always-On continuous collection of voice, in-cabin vision, and vehicle body sensor signals, sending user requests and environmental information to the Sage model for inference → decision-making → receiving the task plan output by Sage, and can flexibly call over 100 in-vehicle tools (air conditioning, navigation, seats, audio and video, etc.) across the whole vehicle for execution.
Image Source: SenseAuto
Large Language Models (LLMs) learn massive amounts of textual knowledge through pre-training and excel at writing, summarization, and language reasoning. However, their performance plummets when encountering non-linguistic, symbolic, and spatial unfamiliar tasks (OOD, out-of-distribution tasks). Behind this lie two major contradictions. First, there is a mismatch between the action space and the generation space; that is, generating each token for an LLM requires running massive parameters. Letting the large model explore the environment through trial and error on its own results in extremely high computing costs and extremely low efficiency. Second, pre-training texts are difficult to completely encode real-world physical/environmental rules. Book texts cannot exhaust all dynamic environments, making it very strenuous for the LLM to explore new environments purely on its own. In in-vehicle scenarios, it is costly and slow to converge for edge-side large models to learn cockpit-driving linkage and complex driving scenario planning from scratch through trial and error.
Sage is positioned as an edge-side AI large model with high-level Agent reasoning capabilities on the edge side. It adopts an MoE architecture with a total parameter count of 32B and only 3B activated parameters. Its differentiated advantages are zero token cost and Always-On offline availability. In offline scenarios, all in-vehicle cockpit agent tasks are independently completed by the local Sage 3B edge-side model in the vehicle, without the need to upload problems to remote servers for inference, and there is no token billing. The underlying working mechanism is realized through two key technologies independently developed by SenseAuto around the post-training stage of Sage: SCOUT and ERL.
Differentiated Advantages of SAGE
Image Source: SenseAuto
Deconstruction of SAGE Key Technologies: SCOUT + ERL
SCOUT (Sub-Scale Collaboration On Unseen Tasks) is a large model agent training framework jointly published by SenseAuto and Nanyang Technological University at ICML 2026, and it is also the core post-training technology for the Sage edge-side large model. The framework adopts a three-stage pipeline:
Stage 1: Exploration Stage
Using lightweight small networks (small MLP/CNN) as scout soldiers, the small model has a small number of parameters and fast inference speed;
Leveraging reinforcement learning (DQN/PPO) to quickly try and error in unfamiliar environments, producing a large amount of high-quality expert trajectories (action sequences, environmental interaction experiences);
Stage 2: Distillation Stage
Textualizing the interaction trajectories run by the small model, and distilling environmental knowledge to the large model through supervised fine-tuning. The large model does not need to try and error from scratch, but directly learns the environmental rules explored by the small model.
Three-Stage Framework of SenseAuto SCOUT Large Model Agent Training
Image Source: SenseAuto
Stage 3: Evolving Stage | Multi-Round Reinforcement Learning Activates Intrinsic Knowledge of Large Models
On the basis of learning environmental rules, multi-round PPO reinforcement learning is performed on the LLM to activate the world knowledge brought by the pre-training of the large model. Combined with new task rules, it optimizes reasoning and decision-making capabilities. This solution solves the pain points of high computing costs and slow convergence when pure large models autonomously explore unknown scenarios. It can reduce training computing overhead and improve the model's reasoning capabilities in spatial, long-term sequential, and open-ended unfamiliar tasks.
Deconstruction of the Working Principle of ERL, One of the Key Post-Training Technologies of SenseAuto Sage Model
Image Source: SenseAuto
ERL (Erasable Reinforcement Learning) specifically solves the core problem of "one step wrong, the entire task chain fails" in multi-step chain reasoning for search-augmented large models. The mechanism draws on human thinking and error-correction habits: the model can autonomously locate the error links in the reasoning chain, only erase the erroneous fragments, retain the verified correct historical information, and only locally retry planning for the erroneous parts, without having to start all tasks from scratch, effectively improving the success rate of small models in executing complex continuous tasks.
For example, suppose a user issues a long-chain command: "Navigate to Hangzhou, find a favorite restaurant along the way, and reserve a parking space." For such multi-step Agent chain tasks, the working steps of the edge-side AI model are: decompose the task → query location → query restaurant → query parking space → issue navigation. Using traditional AI large models, if there is an error in decomposition or understanding at any intermediate step, the entire process will directly collapse. SenseAuto's ERL allows Sage to identify intermediate step errors during the reasoning process, locally erase the erroneous links, and re-plan without having to start all tasks from scratch, thereby improving the success rate of complex in-vehicle Agent tasks.
From reducing the learning cost of unfamiliar environments at the source of training to repairing chain task errors during the inference operation stage, SCOUT and ERL complement each other and jointly build the algorithm foundation of the Sage edge-side large model for in-vehicle embodied intelligence. On the basis of this technology, SageBox, implemented in the form of an external AI-BOX, also reflects the current opportunities and realistic constraints of the entire Physical AI industry. This edge-side large model agent capability is not limited to automotive scenarios; it can also be transferred and reused to serve as the edge intelligent brain for other hardware, expanding to Physical AI terminals such as home edge brains and companion robots, releasing the value of embodied intelligence on more physical devices.
Key AI BOX Manufacturer 2: Mass-Produced AIBOX-N1 by ThunderSoft
The mass-produced AIBOX-N1 by ThunderSoft is an automotive-grade independent AI computing power expansion unit built on NVIDIA Drive AGX Orin. It stably outputs 200 TOPS of AI computing power and 205 GB/s ultra-high bandwidth, perfectly adapting to the smooth local inference of 1–7B multi-modal large models. The core advantages of the product are lightweight external deployment and zero whole vehicle architecture modification. It supports dual cooling solutions of air cooling and water cooling, adapting to the installation environments of various passenger vehicles and wide-temperature automotive-grade working conditions.
Relying on the self-developed Drip AIOS 2.0 Pre system foundation, the product can achieve multi-model concurrent scheduling, dynamic computing power allocation, and low-latency inference optimization, effectively solving the pain points of computing power congestion in traditional cockpit domain controllers and the lag in large model operation. It helps automakers quickly implement high-level functions such as cockpit AI Agents, multi-modal interaction, enhanced sentry, and proactive scenario services without replacing the original vehicle domain controller. It supports layered configuration of computing power for vehicle models, greatly shortening the intelligent facelift cycle and reducing R&D iteration costs. Currently, this product has obtained pre-installation design wins from leading OEMs such as Geely, and is a benchmark flexible computing power solution for the large-scale implementation of in-vehicle edge-side large models in the mid-term of the industry.
Mass-Produced AIBOX-N1 by ThunderSoft
Image Source: ThunderSoft
Key AI BOX Manufacturer 3: ModelBest AI Large Model
ModelBest AI BOX is an external independent cockpit computing power unit led by large model manufacturers. With the MiniCPM series edge-side large models + SuperMate cockpit AI agent as the core barrier, it provides OEMs with an integrated solution of "hardware computing power foundation + native edge-side large model + cockpit Agent application". Without the need to reconstruct the original vehicle EE architecture, it provides flexible computing power expansion for existing facelifts and brand-new models, prioritizing the realization of a purely local offline closed-loop multi-modal cockpit intelligent experience.
The product adopts a dual-platform layered layout to accurately cover the high-end and low-end markets: the flagship version is jointly built with the Intel Core Ultra platform, equipped with an XPU heterogeneous architecture, and provides up to 180 TOPS of dense AI computing power. It can support 35B ultra-large parameter models and complex long-context, full-modal reasoning tasks, adapting to high-end intelligent models. The cost-effective version adopts the Rockchip RK3576M + RK1828 automotive-grade dual-chip solution. Relying on air cooling and low power consumption characteristics, it stably runs 3B–7B MiniCPM models with an inference speed of up to 100+ tokens/s, adapting to the large-scale mass production of mid-end models.
ModelBest AI BOX Adopts Dual Hardware Platform Layout
Source: Zosauto
Facing the second half of Physical AI, the evolution of automobiles into super intelligent agents on wheels has become a clear industry trend. Achieving proactive full-process services and cross-domain collaboration among cockpit, autonomous driving, and ecosystems is the core goal of in-vehicle embodied intelligence, but the implementation of ideal capabilities still faces realistic constraints. As an incremental upgrade solution, the external edge-side AI BOX can make up for the computing power of large models without drastic changes to the whole vehicle architecture. It has become a pragmatic path that balances cost and iteration efficiency at this stage, and has also attracted multiple players including large model enterprises, Tier 1s, chip manufacturers, and OEMs to enter the competition.
However, the external AI BOX is still a phased transition route. In the future, with the successive mass production of new-generation cockpit-driving integrated chips, the native integration of large models in domain controllers will become an important development direction. Although the hardware form will iterate, the essence of industry competition always comes down to the real user experience. How to control hardware costs, meet automotive-grade safety, optimize inference latency, and turn demonstration-level long-chain proactive services into high-frequency available functions is a common challenge for the industry.