Gasgoo Munich- By 2026, the buzz around embodied intelligence hasn't faded, but the debate has shifted: Is VLA truly obsolete?
For the past two years, VLA (Vision-Language-Action models) were practically the industry standard for embodied intelligence foundations: take in vision and language, spit out actions. This end-to-end architecture—perception, decision, execution—powered the sector's first leap from simple repetitive demonstrations to generalized instruction following.
But in 2026, "world models" have become the new hotspot—specifically, the World Action Model (WAM) for robot control, which is moving rapidly from academic papers to industrial practice. As more teams commit to this approach, a quiet mantra is making the rounds: "VLA is dead; long live WAM."
Amid the debate, some firms are pivoting entirely to WAM, others are holding the line on VLA, and a few are hedging their bets on both. A new industry consensus is taking shape: VLA is far from finished, and WAM is no magic bullet. The race to build the foundation model for physical AI is nowhere near an architectural endgame.
Clarifying the Core: The Essential Divide Between Two Paths
To understand the rivalry between VLA and world models, we have to look at their fundamental differences. Put simply, the core distinction lies in their logic for getting things done.
VLA operates much like human "muscle memory": see the scene, hear the task, and the brain immediately commands the body to act.
More specifically, VLA is a direct mapping from perception to action. It takes visual images and language instructions, runs a calculation, and outputs the next move—no extra "thinking" or simulation required.
The advantages are straightforward. "VLA excels at semantic understanding, task grounding, and coarse-grained task planning," says Yao Maoqing, partner, senior VP, and president of embodied intelligence at Agibot, and chairman and CEO of Mifeng Technology. "In short, it helps us clarify what the object is, what the environmental state is, and what task we need to complete."
Image Source: Agibot
Beyond that, VLA offers high inference efficiency. Without intermediate simulation steps, the computational load is manageable, making it easier to meet the low-latency demands of real-time robot control.
But the weaknesses are just as obvious. By skipping cognitive modeling of the world environment, VLA struggles to adapt when facing new scenes or tasks not covered in training data, leading to a noticeable drop in generalization capabilities.
"We've found that traditional VLA models are prone to performance degradation—that is, their capabilities decline when encountering new scenarios or tasks," Su Hang, an associate researcher in the Department of Computer Science at Tsinghua University, said bluntly at the "Embodied Intelligence: Embracing the Emergence of Physical AI" forum during WAIC 2026.
Moreover, lacking an internal causal model of how the physical world works, VLA can't simulate the consequences of different actions before moving, the way humans do. This makes it difficult to handle complex tasks that demand high levels of safety and planning.
WAM takes the opposite approach, resembling the human habit of "thinking before doing." Before lifting a finger, it simulates the process and result of the action in its mind, executing only once it's sure. In other words, WAM doesn't output actions directly from the current visual. Instead, it generates a visual prediction of what the world will look like after the action, then reverse-engineers the necessary moves to achieve that goal.
This "simulation-based" architecture offers three natural advantages. First, stronger long-horizon task planning. By planning coarse-grained intent over a future timeline while executing fine-grained physical control in the immediate moment, it achieves "precision up close, rougher at a distance"—avoiding the myopic trap of taking it one step at a time and drifting off course.
Second, better robustness. Because WAM anticipates the world state after an action during inference, this "predict-verify" mechanism provides a theoretical basis for anomaly detection and strategy adjustment. If reality diverges from the prediction, it can actively correct its course, rather than getting stuck in the "dead loops" that plague VLA.
Third, higher utilization of video data. WAM's predictive modeling logic aligns naturally with massive amounts of internet video, allowing it to learn the laws of physics more efficiently and potentially reducing reliance on expensive real-world demonstration data.
Of course, WAM isn't perfect. Its semantic understanding is weaker than VLA's, making it less responsive to complex natural language instructions. And because it must simulate future physical states before acting, the computational cost of inference is higher, placing greater demands on the computing power and bandwidth of robot-side chips.
Because each has strengths and weaknesses, VLA and WAM aren't really substitutes in the same dimension. Technically, they correspond to two distinct core capabilities of embodied intelligence: one leans toward cognitive and semantic understanding, the other toward physical and motion control. Both have clear boundaries.
Choosing a Path: The Inevitable Trend of Convergence Through Scenario Adaptation
The debate isn't over, but few companies are sticking to a pure "VLA-only" or "WAM-only" dogma. A new consensus is forming: these paths will likely converge. The question is how, how fast, and who will lead—and opinions vary.
Guo Yandong, founder and CEO of AI² Robotics, has argued on multiple occasions that VLA and world models aren't replacements but deeply integrated, symbiotic partners. "The human brain has a built-in world model to understand the physical environment and anticipate changes, while relying on high-level cognition for task planning—that corresponds to VLA capabilities in robots. The structure of the human brain reveals the technological endgame: fusion is the only answer; there is no choosing one over the other."
At the forum, Yao Maoqing also stated clearly that the endgame for both paths is fusion. Accordingly, Agibot proposed the World Reasoning Action Model (WRAM). "WRAM works by first understanding the task, then simulating the future, evaluating options, and finally executing the action," Yao explained. "It's no longer a simple perception-action muscle mapping, but an ultimate agent capable of simulating the physical world."
A robot deploying the Act2Goal model practices drawing unseen patterns. Image Source: Agibot
To achieve this, Agibot has built a three-stage training architecture: pre-training, post-training, and continuous learning. It is evolving along both the VLA and WAM tracks.
On the VLA front, Agibot has released two foundation models, GO-1 and GO-2. GO-1 gave robots reliable understanding, while GO-2 focused on bridging the gap between planning and execution, introducing a chain of thought for actions for the first time. Building on this, Agibot expects to launch the GO-3 model by the end of the year.
On the world model front, Agibot has launched GE-Act, which supports memory, and Act2Goal, which achieves Goal-to-Action for the first time, marking a leap from future prediction to closed-loop robot control.
Ren Zhiyi, a research scientist at Physical Intelligence (π), agrees that there is no conflict between VLA and WAM.
As a top Silicon Valley embodied AI startup, Pi positions itself as building a general-purpose embodied brain. Its core goal is to maximize generalization and steerability so one model can adapt to different bodies, scenarios, and tasks. To that end, PI has released several generations of embodied brain models over the past two years, with the latest, π0.7, launching in April.
π0.7's core approach adds a future visual sub-goal generation module to a mature VLA architecture. It uses a world model to generate a "target image for the next step," then uses VLA to output specific actions based on that target. This preserves VLA's advantages in semantic alignment and generalization while patching the shortcomings in long-horizon planning and state prediction—all with lower engineering complexity.
"So far, I haven't seen any demo more impressive than π0.7's, so VLA probably isn't dead yet," Ren Zhiyi said bluntly, responding to the "VLA is dead" narrative.
Dyna Robotics represents the pivot to WAM.
As a representative startup focused on overseas deployment, Dyna Robotics' core goal isn't to build the most general model, but to maximize success rates in specific scenarios—keeping robots working 24/7.
Ma Yecheng, co-founder of Dyna Robotics, points out that their shift from VLA to WAM is driven entirely by deployment needs.
Still, Ma admits that both have their merits; neither is absolutely better. It depends on the scenario. "For the different capabilities we need, some scenarios are more efficient with WM, others with VLA."
Tencent has skipped the "which single model is better" debate entirely, opting for a layered architecture.
Image Source: Tencent Robotics X Lab
Zhang Zhengyou, chief scientist at Tencent and director of the Robotics X Lab, believes embodied intelligence is still in its early stages without a unified framework. Unlike large language models, where Transformer is settled and progress comes from adding data and compute, here things are still fluid.
So Tencent built Apexio, a continuously online embodied agent with three layers: a cognitive system for reasoning and planning at the top, a perception-action system for decision-making in the middle, and a physical reaction system for real-time motion control at the bottom.
Each layer operates in a closed loop with shared information, running at different frequencies. In this system, VLA and WAM are just tools at different levels—neither replaces the other.
Notably, alongside Apexio, Tencent recently released three major solutions: the new embodied VLM foundation model Hy-Embodied-VLM-1.0, the embodied-native world cognitive model Hy-Embodied-RxBrain-1.0, and the Vision-Language-Action (VLA) model Hy-Embodied-VLA-0.5. All have entered real industrial scenarios, completing deployment verification in tasks like shopping guidance and elderly care.
Overall, while corporate paths seem divergent, they point to a clear industry consensus: at this stage, no single architecture can cover every scenario required by physical AI. The fusion of VLA and world models is the inevitable direction of technological evolution.
Endgame Speculation: Long-Term Evolution with Parallel Paths
Behind the continuous evolution of embodied models, the industry is asking one question: Will embodied intelligence converge on a unified architecture like large language models, entering a phase where scaling laws—adding data and compute—dictate progress?
For now, at least in the short term, that seems unlikely.
The reasons aren't complex. First, the physical world is far more complex than the digital one.
Language is one-dimensional, discrete, and rule-bound. The physical world is three-dimensional, continuous, and filled with uncertainty—friction, deformation, collision, disturbance. The variables are endless. Relying on a single architecture to cover semantic understanding, physical simulation, and motion control all at once is a tall order.
Second, the constraints of edge computing make a layered approach inevitable.
Yao Maoqing did the math: even the most advanced edge chips can only run models with 5 billion to 10 billion parameters, at roughly 10 Hertz. You can't put a model with hundreds of billions of parameters directly on a robot body to control motors moving thousands of times per second.
This forces the industry toward a hybrid cloud-edge, layered control path: large cloud models for low-frequency cognitive planning, small edge models for high-frequency action execution. In this architecture, a single "do-it-all" model simply doesn't exist.
Finally, the fragmentation of scenarios dictates a diversity of technical routes.
Industrial manufacturing, home services, warehousing and logistics—different scenarios demand vastly different levels of precision, speed, cost, and generalization. No single architecture can be optimal everywhere.
Because the era of a unified architecture is still far off, the embodied intelligence track has formed a pattern of parallel development across diverse technical routes. Beyond the mainstream paths of VLA, world models, and their fusion, several differentiated paths are advancing simultaneously.
Image Source: EBKernel
Brain-inspired intelligence is one such path. Representative domestic companies include EBKernel, AI² Robotics, QJ Robots, and LivSyn Robotics.
This path avoids the traditional "big data + big model" fitting approach. Instead, it advocates replacing blind computation with clear brain-inspired principles, replicating human cognitive logic to build a machine mind capable of abstract reasoning, autonomous planning, and continuous learning.
In the view of Zhu Senhua, founder and CEO of EBKernel, Embodied Intelligence 1.0, represented by the VLA system, has hit its ceiling. Global technical routes are collectively shifting toward world models, marking the industry's entry into Embodied Intelligence 2.0—characterized by simulation, planning, and autonomous decision-making. However, the industry's definition of "world model" remains unclear, and the direction of algorithm evolution is uncertain.
For these reasons, EBKernel believes that breaking through the limits of embodied intelligence won't come from merely accumulating data, but from revolutionary innovation in the underlying paradigms of neural networks. Specifically, brain-inspired intelligence is the only core path to a breakthrough—and the necessary route to AGI.
To this end, EBKernel has developed two brain-inspired embodied technology foundations: Bi-VLA (Brain-inspired VLA) and Bi-Agent (Brain-inspired Agent). Using these, it built the first generation of its brain-inspired embodied brain product, EBKernel Brain. This solution supports "one brain, multiple machines" and "one brain, multiple forms," and has already undergone deployment verification.
Guo Yandong also believes that "brain-inspired" capabilities will be a key element in breaking through the next generation of physical world AI.
In his view, embodied large models aren't just a game of stacking compute and data. Even combining world models with VLA falls far short of a final solution for a robot brain. "To truly solve the core challenges of training efficiency, computing power, and energy consumption in the future, the key lies in brain-inspired technology."
Image Source: AI² Robotics
Focusing on this goal, AI² Robotics adopted a "cerebral cortex-cerebellum-spinal cord" three-layer architecture to build NeuroVLA, the first brain-inspired embodied intelligence system, launched in June. In this system, the cerebral cortex handles semantic understanding and task planning, the cerebellum manages high-frequency motion coordination and dynamic correction, and the spinal cord handles millisecond-level execution and safety reflexes.
The system is currently deployed on AI² Robotics's AlphaBot 2 robot, which was unveiled during WAIC 2026. It demonstrated capabilities in industrial manufacturing, specifically PCB loading and unloading, completing a series of continuous actions like tray picking, transport, and dual-arm coordination.
Notably, this pattern of parallel diverse paths is itself a signal that the track is maturing.
Two years ago, the industry was still debating the industrial value of embodied intelligence. A year ago, companies were competing on who had the most impressive demo videos. Today, practitioners are discussing the specific scenarios, cost efficiency, and integration paths of different technical routes. This means the industry has moved past the conceptual enlightenment phase and entered a stage of solid technical deepening and deployment verification.
Conclusion
The laws of the digital world can be fitted with massive data, but the physical world is full of complexity and uncertainty. This dictates that the evolution of embodied intelligence will not have just one standard answer.
So, there's no rush to declare VLA obsolete, nor to hype WAM as a new myth. At least for now, VLA's accumulated strengths in semantic understanding still hold value, and WAM's breakthroughs in physical simulation are progressing. Combined with explorations into differentiated paths like brain-inspired intelligence, these efforts will collectively broaden the industry's technical boundaries.
Ultimately, the value of all technical routes must be tested by real-world deployment. When robots can stably and cost-effectively solve concrete problems in the real world, the debate over which route is best will naturally settle itself.
