The embodied-AI industry is becoming more practical. A robot that walks, gestures, or performs one carefully prepared demonstration can attr...
The embodied-AI industry is becoming more practical.
A robot that walks, gestures, or performs one carefully prepared demonstration can attract attention. The harder question is whether it can work for months in restaurants, hotels, factories, warehouses, shopping centers, and other unpredictable environments—and become more capable through that work.
Pudu Robotics used WAIC 2026 to frame its answer around a simple strategic idea:
The next stage of robotics will be decided not only by the body of a robot, but by the intelligence system that can learn from real deployments and transfer that experience across many bodies.
The company calls this architecture “One Brain, Multiple Embodiments.”
Its foundation-model layer, PuduFM 1.0, is intended to help robots reason about physical consequences, connect navigation with manipulation, and learn through a combination of simulation and real-world feedback.
Its agent platform, PuduAgent, packages planning, memory, execution, reusable skills, and safety controls into a system that can coordinate different robot forms.
At WAIC, the industrial semi-humanoid PUDU D7 became one of the most visible hardware examples of that strategy.
Pudu entered WAIC with a large installed base in commercial service robotics.
The company says that a Frost & Sullivan independent market study ranked Pudu first in four categories:
Pudu also reports that it has shipped more than 130,000 robots globally. Its official materials describe deployments across more than 80 countries and regions, while its WAIC presentation cited coverage across 85 countries and regions.
These robots operate in sectors including:
Deployment at this scale is commercially meaningful, but Pudu’s embodied-AI argument goes further.
Every robot working in a real location encounters physical variations that are difficult to reproduce completely in a laboratory:
For an embodied-AI company, those interactions can become training material.
The next competitive advantage may therefore come from combining hardware distribution with a system that converts operational experience into reusable intelligence.
Language models can learn from large collections of text, code, images, and video. Robots face a different data problem.
A robot must understand not only what an object is, but also:
These capabilities are difficult to acquire from a limited number of curated demonstrations.
Robots need experience from real work, but getting into real work requires a level of maturity that is itself difficult to achieve without real data.
That creates a familiar chicken-and-egg problem:
Without large-scale deployment, a company lacks diverse physical-world data. Without diverse physical-world data, it is harder to improve the robot enough for large-scale deployment.
Commercial service robots provide one way to break this cycle.
Pudu spent years placing delivery, cleaning, and industrial robots into routine business environments. Those systems did not begin as general-purpose embodied agents, but their deployment created a large channel for collecting operational signals.
Pudu says its fleet now produces annually:
| Data category | Company-reported annual volume |
|---|---|
| Real-world navigation data | 36.5 million hours |
| Real-world manipulation data | 15.8 million hours |
The company says its manipulation data is collected through an “unobtrusive data collection” mode during normal operation.
The value is not only the number of hours. Much of this material comes from the robot’s own perspective and can include real depth, scale, motion, lighting, and interaction conditions.
That first-person physical context is difficult to replace with internet video alone.
Pudu’s long-term model can be described as a data flywheel:
This is different from treating a robot as a finished appliance.
In the traditional model, a company sells hardware with a fixed set of capabilities. Each completed task creates value for the customer, but the experience may remain local to that device and that installation.
In a physical-agent model, an individual task can also contribute to a broader learning system. A navigation failure in a warehouse, a difficult cleaning condition in a shopping center, or a manipulation correction in an industrial setting may help improve future behavior across the fleet.
That does not happen automatically.
Raw logs are not the same as useful model context. The data must be:
A large fleet creates the opportunity for a data moat. The quality of the data pipeline determines whether that opportunity becomes a real capability advantage.
Physical mistakes are expensive.
A language model can generate a wrong paragraph and try again almost immediately. A robot’s failed action can consume time, damage equipment, interrupt business operations, or create a safety risk.
Real-world training therefore cannot scale through unrestricted trial and error.
Pudu’s answer is a dual-space evolution flywheel that combines simulation and reality.
A high-fidelity world simulator allows the system to generate and test large numbers of trajectories without exposing real people or equipment to every failure.
The simulator can create:
Pudu says it uses a Diffusion Transformer architecture to generate diverse trajectories in the simulated environment.
The real environment supplies the details that simulation cannot fully reproduce.
Human-in-the-loop feedback can record:
The two loops serve different purposes:
The effectiveness of this method depends on simulation fidelity and the quality of the transfer between virtual and physical systems. A model that succeeds in simulation can still fail in the real world because of sensor noise, mechanical wear, imperfect geometry, latency, or unmodeled human behavior.
Historically, different robot forms were often developed as separate systems.
A delivery robot, cleaning machine, industrial mobile base, robot arm, humanoid, and quadruped may use different sensors, joints, control frequencies, payloads, and operating constraints.
This led to an inefficient pattern:
When the hardware changed, much of the work started again.
Teams rebuilt environments, recorded demonstrations, trained specialized policies, and wrote new integration logic. Data collected for one robot was difficult to reuse on another.
Pudu’s “One Brain, Multiple Embodiments” strategy attempts to move beyond that pattern.
The goal is not to make every body identical. The goal is to create a shared intelligence layer that understands common physical concepts, while leaving each robot’s hardware-specific controller to handle its own joints, sensors, and limits.
The shared concepts may include:
If the architecture works as intended, a lesson learned by one robot form can shorten the adaptation process for another.
Pudu says that, with its world-model training and a small amount of real correction data, a new task can be adapted with around 50 expert demonstrations. This is a company-reported result and should be evaluated per task, robot form, and deployment environment.
The larger point is the shift from memorizing a fixed action sequence to learning physical relationships that can be reused.
PuduFM 1.0 is the foundation-model layer behind the strategy.
Pudu presents it as a continuously evolving embodied-intelligence model designed to address three problems:
The architecture centers on three technical elements.
The Physical Intuition Model, or PIM, is intended to help the robot predict what may happen before it acts.
A human does this constantly.
When picking up a wet glass, a person may automatically use a more careful grip. When moving a tall load, a person considers balance. When stepping onto a slippery surface, a person adjusts speed and force.
A traditional perception system may identify the glass or floor but still lack a model of the physical consequence.
Pudu says PIM uses a causal-attention Transformer architecture and models object motion and physical interaction in latent space.
Rather than concentrating only on dense pixel prediction, it attempts to represent dynamics such as:
The robot’s current position, pose, and proposed action can be evaluated before execution.
In practical terms, PIM is intended to address the problem of a robot that can see an object but does not understand what its planned action will do to it.
The second element is a multimodal Vision-Language-Action system.
Many earlier robot stacks separated:
A task could therefore fail between modules.
The robot might understand the instruction and reach the correct location but fail to turn that understanding into a stable manipulation sequence.
PuduFM’s VLA layer is intended to place language, vision, movement, and manipulation inside a unified representation.
PIM’s physical prediction is then incorporated into the action process.
The resulting sequence is conceptually:
This combined approach is meant to reduce the gap between “getting there” and “doing the job.”
The third element is the world simulator that drives the virtual side of the learning loop.
Pudu says the simulator is constructed using operating data gathered across many commercial environments.
It generates both standard and adversarial trajectories, creating examples that would be expensive, rare, or unsafe to collect repeatedly in the real world.
The world model supports PIM and VLA by providing large numbers of controlled physical scenarios.
The process can be summarized as:
| Stage | Purpose |
|---|---|
| Pre-training | Build broad visual, language, video, and physical commonsense |
| Simulated reinforcement | Practice decisions and actions at scale in virtual environments |
| Real-world fine-tuning | Correct the model using physical deployment and human feedback |
The system is not simply “simulation first, deployment later.”
Pudu describes it as a continuous loop in which real data improves simulation, and simulation helps prepare the model for the next set of physical tasks.
PUDU D7 made its first offline public appearance at WAIC 2026.
Pudu positions D7 as an industrial-grade semi-humanoid robot for environments such as manufacturing, retail, and logistics.
Its wheeled lower body is designed for stable and efficient movement, while the upper body provides two arms for manipulation.
This form reflects a practical compromise.
A fully bipedal humanoid may be useful where stairs and human-like mobility are essential, but wheels can offer better efficiency and stability on the flat floors common in warehouses, stores, and factories.
Pudu’s official WAIC materials say D7 supports payloads of up to 14 kilograms and is designed for extended industrial operation.
The strategic importance of D7 is not only its body.
It is presented as one of several embodiments that can use the same broader PuduFM and PuduAgent intelligence stack.
Describe your idea once, and We0 AI can generate a showcase site, pages, and CMS, then help you attract customers and traffic after launch.
One complete project generation for free registration
Best for trying one complete generation flow and seeing a first project draft quickly.
That makes it a test of whether experience can move across Pudu’s existing service, cleaning, and industrial fleet into more general manipulation systems.
A robot foundation model becomes more valuable when it can support more than one machine.
Pudu’s portfolio includes four broad categories:
Within those categories are specialized mobile robots, semi-humanoids, humanoids, and quadruped systems.
The “One Brain, Multiple Embodiments” architecture separates the system into three layers:
Physical Agent = Embodiment + System + Skills
The embodiment is the physical machine:
Different bodies remain optimized for different jobs.
The system coordinates the robot over time:
This is the role Pudu assigns to PuduAgent.
Skills are reusable capabilities such as:
PuduFM provides the shared physical-intelligence foundation behind these capabilities, while PuduAgent turns them into operational workflows.
If PuduFM is the shared physical brain, PuduAgent is the layer that manages how that intelligence is used.
Pudu describes PuduAgent as a general embodied-agent platform for the physical world.
It contains three major parts:
The operating layer contains three core components.
Agent Core interprets the current environment and task, then plans what the robot should do.
Its responsibilities may include:
Agent Memory records useful information from the task process.
For a physical robot, memory involves more than saving text.
It may need to preserve:
Persistent memory helps the robot avoid repeating the same mistake and maintain its goal across a longer operation.
Agent Runner turns high-level decisions into real-time execution.
Planning can happen at a relatively abstract level, but motors and sensors operate continuously and require low-latency control.
The runner acts as the bridge between the plan and the robot’s hardware-specific execution system.
PuduAgent Skills abstracts common capabilities into a standardized library.
Without this layer, each robot form may require a separate implementation for functions that are conceptually similar.
A standardized skill can describe the intent and interface of an action while allowing each embodiment to execute it according to its own hardware.
Examples include:
This does not mean that one identical low-level controller runs on every robot.
A humanoid arm and a mobile delivery base still require different control logic. The reusable layer is the task definition, planning interface, safety contract, and learned physical experience.
Safety becomes more important as robots operate for longer periods with less direct supervision.
Before executing an action, the system may need to determine:
PuduAgent Safety is presented as a guardrail layer that predicts potential risk and adjusts or blocks actions when necessary.
This is essential for commercial deployment.
A robot does not create value by completing one impressive demonstration. It creates value by performing repetitive work reliably while keeping failure rates, interventions, and safety incidents within acceptable limits.
At first glance, the robotics market appears to be a competition over hardware:
Those factors still matter.
Pudu’s WAIC message suggests that the next layer of competition will also involve four less visible assets.
Companies with large fleets can observe a wider range of environments and edge cases.
Unstructured logs do not automatically improve a model. Data quality, privacy, labeling, and evaluation determine whether the experience is useful.
A learning system becomes more scalable when experience from one body or task can reduce the work required for another.
A foundation model alone is not enough. Robots need memory, execution, skills, safety, scheduling, and operational controls.
The company that integrates these layers effectively may be able to turn every deployment into an input for the next generation of products.
Pudu’s strategy is supported by significant commercial deployment and new financing.
In April 2026, the company announced that it had raised nearly $150 million**, bringing its valuation above **$1.5 billion and cumulative funding above $300 million.
At the time of that announcement, Pudu reported more than 120,000 units shipped and a 23% share of the global commercial service robotics market based on an earlier Frost & Sullivan study.
By the PuduFM and WAIC announcements, the company was reporting more than 130,000 units.
The funding can support work across:
Capital and shipment volume do not guarantee that a general physical-agent architecture will succeed.
They do give Pudu two important resources: a large deployment base and the ability to invest in the systems required to learn from it.
The “One Brain, Multiple Embodiments” direction is compelling, but several difficult questions remain.
A lesson from a cleaning robot may improve shared navigation or obstacle understanding, but it may not directly teach a humanoid arm how to manipulate a deformable object.
Transfer quality must be evaluated by skill, environment, and embodiment.
A large fleet can produce enormous amounts of repetitive, noisy, incomplete, or biased data.
The useful metric is not total hours alone. It is how much validated and diverse training signal those hours produce.
Robots operating in public and commercial environments may capture sensitive visual, spatial, or operational information.
Data governance must include consent, retention, access control, anonymization, regional regulation, and cybersecurity.
A world model can expand training, but the simulation-to-reality gap remains one of robotics’ central challenges.
Models must be tested against physical hardware under real operating conditions.
As robots handle longer and more complex tasks, the number of possible failure paths increases.
Safety systems need deterministic controls, hardware protections, monitoring, rollback, and human escalation—not only model-level reasoning.
Customers ultimately care about total operational value:
A technically advanced robot must still outperform simpler automation on the economics of a real task.
When robots begin to work at scale, the most valuable output may not be the number of completed deliveries, cleaning routes, or transport jobs.
It may be the reusable physical experience created by those tasks.
Pudu’s strategy links four layers:
PUDU D7 brings that strategy into a more general industrial robot form, but the larger bet is architectural.
Pudu is trying to turn a product portfolio into a learning ecosystem.
The long-term competition will not be won solely by the robot that performs the most impressive isolated action. It will be shaped by which system can learn safely from the largest number of useful physical interactions, transfer that learning across machines, and deliver reliable economic value in real environments.
PuduFM 1.0 is Pudu Robotics’ embodied-intelligence foundation model. It combines a Physical Intuition Model, multimodal Vision-Language-Action alignment, and a simulation-to-reality training loop to support navigation, manipulation, and physical reasoning across different robot forms.
PuduAgent is Pudu’s general embodied-agent platform for operating robots in the physical world. It includes an operating layer for planning, memory, and execution, a reusable skill library, and safety controls.
It means that several types of robots share a common intelligence foundation instead of being developed as completely isolated systems. Each body keeps its own hardware-specific control, while higher-level physical understanding, planning, memory, and skills can be reused.
Pudu says it has shipped more than 130,000 robots globally. Its official materials describe service across more than 80 countries and regions, while its WAIC 2026 presentation cited 85 countries and regions.
PUDU D7 is an industrial-grade semi-humanoid robot with a wheeled mobile base and two arms. Pudu positions it for industrial, retail, and logistics tasks and presents it as one hardware embodiment of the PuduFM and PuduAgent architecture.
The Physical Intuition Model, or PIM, is intended to predict the physical result of a proposed robot action. Pudu says it models dynamics such as motion, friction, deformation, balance, and collision risk in latent space before execution.
Simulation allows the system to generate large numbers of normal and adversarial trajectories without repeatedly risking real equipment or people. Real-world operation provides authentic feedback that helps calibrate the simulator and correct the model.
Pudu cites Frost & Sullivan market studies that rank the company first in several commercial-service-robot categories. Those rankings are reported here as the findings of the cited study; the underlying full market dataset was not independently reproduced in this article.
Pudu Robotics used WAIC 2026 to present a strategy that extends beyond individual robot products. Its installed commercial fleet supplies real-world experience, while simulation increases the number and variety of training scenarios.
PuduFM 1.0 is intended to provide shared physical reasoning across robot forms. PuduAgent adds planning, memory, execution, reusable skills, and safety controls. PUDU D7 shows how that stack can be applied to a semi-humanoid industrial platform.
The strategy still faces difficult questions around data quality, transfer across bodies, simulation fidelity, privacy, safety, and commercial return. Its central idea is nevertheless clear.
The next robotics advantage may come from building not just more capable machines, but a shared intelligence system that learns from every machine already at work.
Start from one sentence and have a complete website in minutes.
