Over the past few years, AI's battlefield has been on screens. The GPT family delivered astonishing language ability with sheer scale, and Sora stunned the world with video generation…… But in 2026, the industry has reached a consensus: 2026 is the first year of Physical AI.
At CES in Las Vegas at the start of the year, NVIDIA CEO Jensen Huang delivered a speech and mentioned Physical AI 17 times to declare that 'the ChatGPT moment for Physical AI has arrived.' It is also a keyword he has strongly championed over the past two years. Over the past 2-plus years, Physical AI has made key progress in five dimensions: the 'brain' (VLA models), the 'imagination engine' (world models), training environments, embodiment, and commercial closed loops. The industry now has the technical basis to move beyond demos and explore real-world scenarios.
In the first half of 2026 alone, capital flows in the global Physical AI field showed astonishing density and scale. In Q1, global Physical AI financing exceeded $6.4 billion. This included cases such as AMI Labs' $1.03 billion seed round, World Labs' $1 billion financing, and Qianxun Intelligence's completion of four funding rounds totaling RMB 4.5 billion in March. At the same time, Physical AI's technical routes are showing clear industrial divergence. At the foundation-model layer, VLM, VLA, and world-model paths are converging. World models act as a back-end data factory to synthesize training data, VLM handles the 'slow thinking' of long-horizon planning, and VLA directly turns instructions and perception into actions.
The technical direction seems to have already hardened. The core view is that the future core architecture of Physical AI will be a 'deep closed loop between VLA and world models'—VLA is responsible for 'speaking human language and making decisions,' while world models provide an 'embedded physics engine' that simulates the physical consequences of actions in advance and verifies whether plans are feasible.
But are these solutions already the final answer to the core question of 'what kind of AI does the physical world need'?
Physical AI Is Still an 'Open Competition'
With capital surging in, giants entering the field, and mass-production wins arriving one after another, Physical AI is seen as the next trillion-dollar market.
According to Future Markets, the global Physical AI market will grow from about $383 billion in 2026 to $3.26 trillion in 2040, making it one of the largest technology market expansions in history. Coatue Management is even more aggressive, estimating that the Physical AI market could reach at least $6 trillion, about 50% higher than digital AI.
However, unlike digital AI, where the competitive landscape is becoming increasingly clear, Physical AI is characterized by an 'unfinished structure.' The dimensions that decide victory—manufacturing capability, deployment data, regulatory speed, supply-chain control, and foundation-model intelligence—are spread across different competitors.
As the underlying base of Physical AI, world models are also far from having a consensus on their role. Wang Zhongyuan, president of the Beijing Academy of Artificial Intelligence, has pointed out that there are currently four mainstream world-model paths: the language-centered path (such as Gemini 3), which can perceive multimodal data, think and describe the next state through language, and has planning and decision-making ability; the pixel-centered path (such as Sora), which is suitable for video generation but does not understand physical causality; the 3D-structure-centered path (such as Fei-Fei Li's World Labs Marble), aimed at simulating digital worlds such as the metaverse and games; and the visual-representation-centered path (such as LeCun's V-JEPA series), which predicts compressed visual representations, but the evolution of visual embeddings does not equal the evolution of physical laws.
The four paths each have supporters and each has shortcomings. Within the mainstream VLA and world-model technical camps, some regard VLA and world models as opposites, while others choose to advance both paths in parallel and explore fusion. Some take an embedding route with 'VLA as the main system and world models as support'; others take an independent route with 'world-model native, reinforcement-learning driven'; and still others take a foundational reconstruction route with 'brain-like algorithms + world models.'
Different technical routes are reflected even more sharply at the industrial level.
In the 'VLA route' camp, NVIDIA showcased its in-house VLA model Alpamayo at GTC 2026, along with Isaac GR00T N1.6, an open-source reasoning vision-language-action (VLA) model designed specifically for humanoid robots; XPeng launched X-Foresight at CVPR 2026, embedding a world model directly into the VLA architecture…… In the 'world-model route' camp, Geely Auto Group released the WAM World Action Model at GTC 2026; World Labs continues to bet on spatial intelligence (the 3D-structure-centered world-model route); and Google DeepMind released a general-purpose world model, Genie 3, in August 2025…… In the 'VLA + world-model fusion route' camp, there are also many hot companies, including AgiBot, Unitree Technology, and Tesla Optimus. Among them, Unitree Technology has said it will continue tracking and benchmarking the VLA route while exploring a rational way to integrate 'world-modeling' capabilities into the VLA architecture.
The current lack of a fixed implementation paradigm for Physical AI may be the clearest reflection of the field. On the choice of foundation-model routes, it seems that the three paths of VLM, VLA, and world models are already beginning to converge.
In fact, VLA models break the old limitation of treating vision, language, and action as separate domains. Mainstream VLA models now have stable success rates, and task execution ability is improving quickly. A world model, in essence, is a predictive representation of environmental dynamics. It can inject imagination into VLA, provide an embedded physics engine, and allow AI to simulate the consequences of actions in advance and verify whether plans are feasible.
A purely reactive VLA can only map actions based on current observations and lacks the ability to reason about the future, so long-horizon planning is prone to failure. The world model is the key to breaking through. This is also why the fusion of VLA and world models has become the mainstream trend.
But one question is worth pondering: what should a model for Physical AI actually look like?
The answer may lie neither in a technical route nor in a demo video at a launch event, but in the needs of real-world scenarios. Its core evaluation standard has never been 'is the generated world realistic enough,' but 'can it help machines act better in the physical world,' 'can it reduce trial-and-error costs,' 'can it improve generalization,' and 'can it be embedded into a real business closed loop'……
Physical AI Doesn't Lack 'Hands'
What It Lacks Is the Ability to Read the Physical World
While the industry is busy debating the route war between VLA and world models, a more fundamental issue has been overlooked——for the vast majority of physical devices, what is truly missing is not 'dexterous manipulation' or 'future prediction,' but 'basic understanding of space.'
A video of a robot 'going wrong' in a kitchen once sparked heated discussion on social media. In the clip, a humanoid robot not only knocked over a salad bowl and splashed sauce everywhere, but even lost control and fell to the ground. Yet this was not an isolated case. From flashy performances that can do backflips and dance, to real-world behavior that cannot even fold clothes or clear a table, 'going wrong' is shifting from individual incidents to a common industry problem.
The logic behind this phenomenon is simple: the smooth moves on stage are all the result of preset choreography and repeated debugging, with an absolutely controllable environment; home settings, by contrast, are full of random variables, and unpredictable factors such as clutter and sudden noises are technical challenges for robots. The root cause is that AI's understanding of physical space is far behind hardware capability.
Physical AI is not just humanoid robots and autonomous driving, which are the first things most people think of. It is a vast industrial ecosystem that penetrates the physical world in all directions, with broad application areas covering autonomous robots, autonomous vehicles, human-machine systems, industrial automation, wearables, and AI-enabled medical and agricultural systems. It also includes a large number of vision-capable devices such as cameras, drones, robot dogs, inspection equipment, industrial terminals, and AI glasses.
This reality has drawn the attention of global AI giants, such as Meta's Segment Anything Model 3, Google's Gemma 4 and Vision Banana, and NVIDIA's Vision-FIrst model proposed by Jim Fan. Although these models follow different technical paths, they all point to a common vision: building truly general-purpose visual capabilities.
In fact, for industrial scenarios with fixed environments, fixed coordinates, standardized instructions, and a requirement for tens of thousands of repeated actions with zero deviation, traditional Physical AI is already sufficient. But in scenarios with unstructured clutter, random instructions, and tens of thousands of different environments, what is truly lacking is not a dexterous hand, but continuous perception, spatial understanding, precise localization, and autonomous navigation.
For example, a robot dog does not need a dexterous hand; it needs autonomous navigation in open spaces. A drone also does not need 'hands'; it needs continuous visual positioning in GPS-denied environments. AI glasses do not need to grasp objects; they need precise spatial localization and real-time environmental understanding. Industrial inspection terminals do not need humanoid form; they need stable perception and autonomous decision-making in complex environments……
The mainstream VLA (vision-language-action) and world-model technical routes both have a common blind spot in the dimension of 'spatial understanding.' VLA uses VLM as its backbone, with actions added afterward only as a 'head.' The language part consumes more than 90% of the compute and data, while the action part has long been undertrained. World models try to make up for this shortcoming by building an understanding of physical laws through prediction of the next state of the physical world. But the bottleneck is equally sharp: high-quality data is scarce and far below the scale required by large models.
The essence of a crash is not that the robot is not smart enough, but that it still cannot 'read' the continuous, complex, and uncertain physical world it is in.
This is precisely why Om AI Lianhui has taken a different path over the past five years. It did not chase the hottest direction of 'dexterous manipulation' at the time; instead, it returned to the most basic question in Physical AI: how to make AI first 'understand' the physical world, and only then talk about everything else. The VLX-Flow, VLX-Seek, and VLX-Go series of models released from June 27 to 29 start from the most basic and most universal visual capabilities in Physical AI and propose a new streaming multimodal approach.
Specifically, VLX-Flow solves the 'continuous perception' problem by allowing AI to keep observing like a human instead of passively waiting for questions. Traditional video understanding is 'take one frame, ask once, answer once'——AI watching video is like flipping through a comic book, where every page has to be reopened and reread. Through Linear Attention and dual-layer memory (visual cache + text carryover), Flow lets the video stream 'flow into the model like water,' so AI keeps watching and keeps remembering.
Flow Inference Speed Advantage
VLX-Seek solves the 'precise localization' problem, enabling AI to move from 'roughly seeing' to 'locking on precisely.' VLX-Seek changes localization from 'coordinate generation' to 'region reference'——the former asks the model to 'guess coordinates,' while the latter gives the model a real region to choose from, providing robots and devices with millimeter-level spatial anchors. Different methods, dramatically different results: VLX-Seek can achieve better performance with less data.
Seek Architecture Diagram
VLX-Go solves the 'action decision' problem, taking AI from 'being able to understand' to 'being able to move, follow, and navigate.' It does not output text advice like 'move left'; instead, through the logic of 'short-horizon waypoint prediction + offline trajectory learning + online RL optimization,' it directly outputs executable waypoint trajectories so the device can move by itself.
The key assumption behind this paradigm shift is that AI in the physical world should not be a passive system that only responds when users ask, but an active one that keeps watching, keeps remembering, and can move at any time. The former is the interaction logic of digital AI——humans initiate the conversation; the latter is the working logic of Physical AI——changes in the environment itself are the trigger signal.
In this paradigm, visual information does not enter the model as 'one captured frame' but as a 'continuous stream.' The model does not 'look first and then speak'; it 'understands while watching and acts proactively when needed.' What this corresponds to is not a 'better human-machine dialogue experience,' but a qualitative leap in AI's autonomous work capability.
The significance of this shift may not be that it makes the model 'better,' but that it fundamentally changes how AI exists in the physical world——from 'waiting for the user to ask' to 'working online all the time.'
When hundreds of millions of cameras, drones, robot dogs, and wearables are no longer just 'recording tools' but become 'perception terminals' that continuously understand the physical environment and autonomously complete the full loop of perception, localization, and decision-making——the industrial value of this transformation may be far greater than a stunning robot demo.
Bringing Physical AI into Industry
Physical AI has never lacked narratives about 'technical height,' but the real turning point for industrialization is whether the technology can create quantifiable business value under real constraints——that determines how high and how far it can go.
The VLX series includes many technical innovations, such as replacing standard Attention with Linear Attention, using a dual-layer memory mechanism, replacing coordinate generation with a region-reference paradigm (Region Token), and replacing long-horizon planning with short-horizon waypoint design. These innovations offer an observation sample: when a technical architecture moves from the lab into the real physical world, what real problem does it actually solve?
First, on the edge: in scenarios such as industrial inspection and emergency management, network disconnection is the norm rather than the exception. A drone flying in a factory zone, or a robot patrolling a subway construction site…… if every decision has to wait for a cloud round trip, any delay means 'too late.' Edge deployment completes the decision loop on the device itself, compressing latency from seconds to under 0.1 second——this is not an experience issue, but a fundamental question of whether the business can exist at all. On-device inference compresses the decision loop to the device locally, reducing latency from seconds to under 0.1 second. In industrial practice, this directly moves many high-risk, highly dynamic scenarios from 'technically infeasible' into the range of 'commercially deployable.'
Next, the native architecture: the decision path of VLX-Go is completely defined by its application scenario——it is precisely adapted to lighter short-horizon waypoints so it can respond more quickly to environmental changes. At the 0.6B parameter scale, it does not do 'deep reasoning'; it only does 'real-time response.' This is not a technical compromise, but a clear-eyed understanding of the capability boundaries of edge devices——industrial drones and inspection robots are often equipped with power-constrained embedded chips, whose compute at best supports real-time inference for models at the scale of several billion parameters. Under this constraint, the value of 'moving faster' is far greater than 'thinking deeper.'
Go Architecture Diagram
More noteworthy is the differentiated R&D paradigm: the industry generally treats edge models as compressed versions of cloud models, porting them down through distillation and pruning; VLX's logic is the reverse——first define the compute ceiling of edge hardware, then design a native architecture that maximizes capability within that boundary. This 'edge-first' R&D philosophy is highly aligned with the industry's shift toward specialization, lightweight design, and hard real-time at the edge, and it is also more in line with Physical AI's rigid demand for deterministic latency.
This model has already been validated in industry. On the robotics side, leading embodied-intelligence companies such as Deep Robotics and Unitree have already completed deployments; on the drone side, low-altitude scenarios such as public security, maritime affairs, natural resources, industrial parks, emergency response, and rescue have already gone commercial; on the security camera side, commercial service integration has already reached the million-camera scale.
Of course, Physical AI is still at an early stage. Whether it is richer human-machine-environment interaction or task execution in complex environments, these are still open problems. But for Physical AI, the greatest impact of VLX is not the breakthrough of a few algorithmic metrics; it is a whole set of reflections on the 'form of intelligent existence'——in the physical world, the yardstick for intelligence is no longer the stacking of parameter counts, but the length of decision latency and the extremity of compute efficiency.
If 2026 is the first year of Physical AI, then it is destined not to be a marathon with a clear route, but more like multi-route climbing in the fog. Some are betting on the deep reasoning of the 'brain,' some are betting on the predictive power of 'imagination,' and some, like Om AI Lianhui, choose a plain but necessary entry point——letting machines continuously understand the world in front of them and make accurate decisions. What the industry should really think about is this: when the final form of Physical AI is still an open question, are we 'using technology to explore scenarios,' or 'using scenarios to reshape technology'? The answer is not in papers, and not in demos, but hidden in the real feedback after every on-site deployment.
