Is intelligent driving entering the era of physical AI? 7 industry guests give their latest judgments

On October 26, the 8th Intelligent Driving Conference and the 2026 Huanxiu Lake Physical AI Industry Development Promotion Conference were held in Xiangcheng, Suzhou.

Beyond world models, reinforcement learning, edge intelligence, virtual simulation, and Robotaxi, different forms of physical intelligence such as robots and low-altitude aircraft began to appear at the same conference. Behind this is a set of new problems that the intelligent driving industry is facing.

On the one hand, advanced assisted driving has entered the stage of large-scale deployment, and Robotaxi has begun to move from technical validation toward operations and commercialization; on the other hand, when AI truly enters physical devices such as cars, robots, and aircraft, how to obtain sufficient data, how to train and validate models, how to reduce development costs, and how to ensure the safety of systems in the real world are becoming new technical challenges.

That day, guests from companies including Momenta, Cao Cao Mobility, and ModelBest, as well as from universities and research institutions, offered their judgments on directions including world models, reinforcement learning, edge large models, virtual simulation, vehicle-road-cloud, and physical AI infrastructure.

Summarizing these remarks, it is not hard to see that intelligent driving is becoming the earliest testing ground for physical AI to enter the real world. And from the autonomous driving of a single vehicle to robots, low-altitude aircraft, and urban transportation systems, different technical routes are answering the same question: how can AI truly move from the digital world into the physical world and form products that can operate at scale? Li Keqiang, Academician of the Chinese Academy of Engineering, Professor at the School of Vehicle and Mobility of Tsinghua University, and Director of the State Key Laboratory of Intelligent Green Vehicle and Mobility

Keywords: intelligent driving safety, vehicle-road-cloud integration, bottlenecks of single-vehicle intelligence

Li Keqiang believes that automotive intelligence has entered the stage of large-scale application and commercial closed loop, and the primary issue at present is the safety of intelligent driving systems. Relying solely on a single vehicle's own sensors and computing power has physical boundaries, and high-level autonomous driving needs to evolve further from single-vehicle intelligence to vehicle-cloud collaboration and vehicle-road-cloud integration.

He summarizes the current bottlenecks of the autonomous driving industry into three categories: first, some product definitions deviate from national and international standards; second, key technologies for high-safety, high-reliability autonomous driving still need to be breakthroughs; third, industry investment and commercial output still do not match.

At the technical level, he believes there are mainly three problems:

Development efficiency is relatively low. At present, a large number of systems still adopt a "chimney-style" architecture, lacking a unified foundation, and there are problems such as high development thresholds, low reuse rates, and long validation cycles.

Li Keqiang proposes that vehicle-road-cloud integration can provide two types of core capabilities: first, establish standardized "data tracks" to provide vehicles with wide-area, multi-dimensional, real-time traffic information; second, form a data service system for AI large models to continuously provide highly reliable data for models.

At present, related solutions have entered industrial validation. Cities including Beijing and Chongqing are carrying out collaborative development with automakers, with application scenarios covering stationary obstacles ahead, "ghost probe" incidents, green-wave passage, and others. Some projects have already entered the mass-production stage; this year, seven automakers have advanced related applications, and China FAW has also announced mass-production adoption.

He believes that high-level autonomous driving needs to solve three problems in the future: face up to the safety bottleneck, break through the physical boundaries of single-vehicle intelligence, and reduce development costs and improve R&D efficiency through vehicle-road-cloud integration. Qu Xiaobo, Academician of the Academia Europaea, Professor at the School of Vehicle and Mobility of Tsinghua University

Traditional digital twins mainly complete the mapping from physical entities to digital entities, but digital-side deduction often takes several seconds, making real-time control of physical systems difficult. Related national key R&D projects hope to focus on solving four problems: achieving digital deduction with generalization capability, compressing deduction time to within 100 milliseconds, establishing an interoperability mechanism between the digital side and the physical side, and reducing end-effector execution errors through cross-scale, multi-level architectures.

Its technical system is divided into three levels:

Component level: improve the design efficiency and execution precision of actuators such as motors. By using physical AI to learn from existing motor data, some single-point calculations can be shortened from hours to milliseconds with an error of about 10%.

Vehicle level: solve the Sim-to-Real Gap, adaptation to different models and dynamics, and rapid validation, applied to autonomous driving, autonomous cruising of aircraft, and autonomous operation of engineering machinery.

System level: achieve collaboration among multiple heterogeneous vehicles, including demand-responsive transit, low-altitude traffic control, and mining engineering machinery.

In terms of flying cars, the team has collaborated with Chery to develop a modular flying car with a compound wing for autonomous cruising, composed of wings, a cockpit, and a chassis, which can be combined respectively into an autonomous cruising aircraft and an autonomous driving car.

In terms of engineering machinery, world models are used to generate candidate trajectories and complete trajectory selection, while optimizing the scooping point and unloading position, enabling engineering machinery to complete continuous operations of environment modeling, scooping, transport, and unloading.

Demand-responsive transit has currently been operating routinely for 18 months in the vocational education city of Zhuzhou, Hunan. There are about 100,000 teachers and students locally, and more than 50% have registered to use it. Compared with taxis, the point-to-point travel time differs by only about 2 minutes, but the average cost per trip is about 2 yuan, compared with about 9.16 yuan for taxis.

In terms of low-altitude traffic, by constructing a holographic environment for takeoff and landing areas, conflict prediction, takeoff and landing sequencing, and resource scheduling are carried out. At present, a general aviation airport in Chongqing has been converted into a flying car takeoff and landing test site. Ren Dongchun, CEO of Tangyuan Technology

Keywords: world model, simulation training, synthetic data, virtual-real closed loop

Ren Dongchun introduced the Yootta WorldLoop world model, a virtual-real integrated closed-loop training and evaluation system for intelligent driving and physical AI.

The entire system covers data collection, data production, algorithm training, algorithm evaluation, real-machine verification, and data feedback, and consists of three platforms: scenario production and editing, algorithm training, and simulation evaluation and real-machine verification.

Among them, in addition to geometric and color information, 3D assets will also be augmented with physical attributes such as mass, friction, elasticity, and material, enabling robots to truly interact with virtual assets.

At present, the related systems already cover five types of scenarios: home, industrial, commercial retail, intelligent driving on the road, and parking.

In embodied intelligence scenarios, URDF models of different robots can be loaded, and tasks such as grasping and navigation can be set. Ren Dongchun stated that if 1,000 grasping data entries are collected through real machines, each taking about 1 minute, the total time would exceed 1,000 minutes; using GPU parallel simulation, data production and training at the same scale can be shortened to several minutes to a dozen or so minutes.

In intelligent driving, onboard or roadside sensors can be used to reconstruct real road environments, and then conditions such as weather, lighting, and the number of vehicles can be changed for training and validation of driving and parking algorithms.

At present, Yootta has opened 500 home scenarios with Shangpin Home Delivery, each containing hundreds to thousands of assets that can interact with robots. In the next stage, it plans to build 5,000 virtual-real corresponding scenarios, release 1 million real assets and 5 million hours of synthetic data, and adapt to about 50 robot hardware bodies. Cao Cao Mobility CEO Gong Xin

Keywords: RoboX, Robotaxi, physical AI Agent, integrated air-ground-space mobility

Cao Cao Mobility therefore proposed the RoboX strategy, hoping to uniformly encapsulate different mobility capabilities such as ride-hailing vehicles, taxis, buses, low-speed logistics vehicles, freight trucks, and eVTOLs into services that can be invoked in real time by AI Agents.

In this system, users in the future may no longer need to separately open a ride-hailing app; instead, AI will automatically select and invoke the corresponding mobility services based on schedules, locations, and personal needs.

Basic mobility services: unified dispatch of transportation tools such as Robotaxis, high-speed rail, and low-altitude aircraft;

Super physical AI Agent: turning vehicles into entry points for real-world services such as tourism, dining, and entertainment;

Real-time three-dimensional data services: using in-vehicle and out-of-vehicle sensors to form real-time environmental data and, under compliance preconditions, serve enterprises and urban management;

Distributed computing power: further utilizing idle computing power of vehicles.

Supporting the RoboX strategy is Robo OS, which is responsible for connecting online AI Agents with offline mobility vehicles and feeding real data such as user preferences, mobility scenarios, fulfillment results, locations, and routes back to AI.

At present, Cao Cao Mobility Robotaxi has covered six cities, with a fleet of more than 200 vehicles; more than 50 Robo One units have been deployed in Changsha and Chengdu, with plans to enter Suzhou. The RAS remote center has been established in Hangzhou and Suzhou.

According to the plan, by 2030, Cao Cao Mobility plans to deploy thousands of Robotaxis and Robo One units in Suzhou, achieving 7×24-hour normalized operations.

Its development path is divided into three stages: in 2026, complete the closed loop of admission and operations for Robotaxi and Robo One passenger and cargo transport; in 2027, begin scale replication as forward-developed Robotaxi models roll off the line; and by 2030, promote the Suzhou model nationwide and to overseas markets. Yang Tong, Chairman of Aisi Artificial Intelligence Technology Co., Ltd.

Yang Tong believes that AI is undergoing three changes: general large models are entering more vertical fields, models are beginning to achieve self-evolution through Skills and training feedback, and large models are beginning to connect directly to physical devices.

Based on this trend, Aishi Artificial Intelligence developed Agnes Harness, hoping to reduce the cost of migrating general models to vertical agents and physical AI applications.

Agnes Harness adopts a front-end and back-end decoupled design, encapsulating model capabilities with capabilities such as MaaS, memory, sandbox, search and tools, context management, and team collaboration, and then adapting to different applications through Plugin and Skill mechanisms.

At the physical AI level, one of its core capabilities is MHS, which, through standard hardware protocols, enables large models to connect in natural language to programmable devices such as microscopes, medical equipment, cameras, and drones.

At present, collaborative experiments between small vehicles and robot dogs have been conducted. The two devices rely on cameras to perceive the environment, and then a large language model directly participates in decision-making, without using traditional remote control and without requiring a complete preset action process.

Another focus is ISI, which mainly includes two parts:

Continuously consolidating capabilities during model operation into Skills;

Through TTT (Test Time Training), enabling the model to continue updating during actual operation.

Yang Tong stated that in traditional models, data collection, training, and inference are completed by different teams, and models often remain unchanged for a long time after going online; TTT hopes to connect data collection, execution feedback, and model updates, enabling models to continuously adapt to specific environments.

At present, related experiments use models with about 0.9B parameters for validation, and some models are updated about every half hour. In game tests, after continuously accumulating Skills, the win rate on some tasks increased from about 3% to 90%, and in three-on-three scenarios from 0 to about 60%.

Currently, the AiShi AI model receives about 65 trillion tokens of monthly API calls, generates about 50 million images per month, and generates about 170 million seconds of video. Agnes-related systems have been used in scenarios such as chip design, operator optimization, drone swarms, and protein synthesis. Gu Gongyao, Vice President of Momenta

Keywords: advanced driver assistance, world model, reinforcement learning

Gu Gongyao believes that China's advanced driver assistance has entered a phase of rapid volume growth, with urban NOA solutions beginning to enter models priced at 100,000 yuan and below, while the global market overall lags China by several years.

Momenta's business model can be summarized as "one flywheel, two legs": one is mass-production driver assistance, and the other is large-scale L4 autonomous driving. Mass-produced vehicles continuously generate real-world road data, which in turn feeds L4 products such as Robotaxi.

At present, passenger cars equipped with Momenta's driver assistance system have exceeded 1.3 million units, with cumulative design wins exceeding 240 and more than 120 models delivered, and cumulative real-world driving mileage of about 20 billion kilometers.

Momenta understands physical AI as "world model + reinforcement learning."

Its R7 world model is mainly divided into three stages:

Pre-training: using tens of billions of kilometers of real-world road data to let the world model learn physical laws, traffic behavior, and road common sense;

Simulation rollout: simulating changes in the behavior of the ego vehicle and other traffic participants;

Reinforcement learning: further training the vehicle-side driving model in the simulation environment constructed by the world model.

Momenta currently uses a unified model architecture to support products from L2++ to L4, with mass-production driver assistance, Robotaxi, and other products sharing the same model and data mainline, setting different redundancy only according to regulatory and safety requirements.

R7 has already been applied to multiple mass-produced vehicles. For example, on complex mountain roads, the system needs to simultaneously handle cyclists, median barriers, and oncoming vehicles to find a brief safe overtaking window; when bypassing illegally parked vehicles in urban areas at night, it also needs to identify distant isolation piers in advance and plan a trajectory to return to the original lane.

Momenta is pre-researching the next-generation R9 world foundation model. The R7 parameter scale is about 1 billion, and R9 is planned to be increased to about 100 billion, while vehicle-side chip computing power is also evolving from the hundreds of TOPS level to the thousands of TOPS level.

R9 hopes to further understand more subtle physical changes, for example, not only identifying mountain rockfalls but also deducing whether the subsequent motion of the falling rocks conforms to the law of free fall. Liu Zhiyuan, co-founder and chief scientist of ModelBest

Liu Zhiyuan believes that the core change in large models over the past eight years is not just the continuous expansion of parameter scale, but the formation of a general technical route that can continuously absorb different types of data.

As models continue to expand, data and computing power are becoming new constraints.

On the data side, starting around 2024, large-scale reinforcement learning has driven models to develop from imitation learning to exploratory learning, enabling models to actively explore and generate new training data.

On the computing power side, if capabilities are still improved only by expanding parameter scale, computing demand will ultimately be difficult to sustain. Therefore, large models also need to improve "intelligence density," that is, using fewer parameters and less computing power to achieve the same level of intelligence.

ModelBest therefore proposed the "density law."

According to its observations, the intelligence density of global open-source large models doubles about every 100 days, meaning that about every 100 days, half the parameters and half the computation can be used to achieve similar intelligence capabilities as before.

Liu Zhiyuan believes this is similar to Moore's Law in the semiconductor industry. In the future, models will continue to become smaller on the one hand, while terminal computing power in phones, cars, robots, and other devices will continue to increase on the other; the two trends will eventually converge, causing large models that could previously only run on servers to gradually enter terminals.

There are still three trends as artificial intelligence enters the physical world:

High intelligence density: reducing the parameters and computing power required for model operation;

Omni-modal perception: understanding the real world through vision, sound, and other sensors;

Autonomous agents: continuously making decisions, learning, and acting in open environments.

ModelBest has already applied on-device models to scenarios such as phones, car cockpits, satellites, and robots, and has carried out cooperation with companies such as Huawei and Samsung.

Calculated according to the density law, if intelligence density continues to double about every 100 days, in less than two years the number of parameters and computing power required for the same capability may drop by two orders of magnitude, compressing part of the capabilities of originally trillion-level models into the tens-of-billions level and deploying them to terminal devices.

Liu Zhiyuan believes that in the future large models will continue to enter terminals such as phones, PCs, cars, robots, wearable devices, and IoT, and on-device intelligence will become an important path for AI to enter the physical world. Leiphone Leiphone