At the 2026 World Artificial Intelligence Conference in Shanghai, six robots took on an unusually demanding public challenge: assembling a l...
At the 2026 World Artificial Intelligence Conference in Shanghai, six robots took on an unusually demanding public challenge: assembling a large Great Wall model from 81,920 miniature building blocks in 15 hours.
The finished structure measured approximately 3.5 meters long, 1.5 meters wide, and 1.1 meters at its highest point. Four desktop robots and two wheeled humanoid robots worked on the installation, handling everything from fine block placement to transporting and assembling larger sections.
According to the original report, the robots operated without remote control or a rigid, prewritten action script. They were expected to perceive the workspace, decide what to do next, coordinate with one another, detect placement errors, and correct mistakes as the build progressed.
One of the robots working beside the large building-block Great Wall display.
The challenge was organized by Dexmal, also known in Chinese as Yuanli Lingji. The company presented it as a public test of whether embodied AI could move beyond short demonstrations and perform a long, precise, physically coordinated task in a real exhibition environment.
Accuracy note: Descriptions such as “world-first,” “fully autonomous,” and “industry-leading” come from Dexmal or the original report. The public demonstration and technical materials support the general setup, but the source does not provide a complete independent audit of every operational detail.
A building-block Great Wall may initially sound like a spectacle designed for an exhibition floor. Broken down into its component requirements, however, the task touches three difficult problems that often prevent robots from moving from laboratory demonstrations into real production work.
The blocks had to connect tightly and consistently. The robots therefore needed to perform controlled movements within a range of roughly 0.1 to 1 millimeter.
That level of precision requires several capabilities to work together:
The challenge was not simply to move objects from one place to another. Each block had to be aligned, pressed into position, and integrated with the surrounding structure without disrupting the pieces that had already been placed.
The original report compares the required stability with the natural tremor of a relaxed human hand, which can be on a similar scale. The practical point is straightforward: the robots had to remain both accurate and consistent over thousands of repeated actions.
The second difficulty was coordination.
The setup included six robots with two different body types. They shared the same physical workspace and had to divide the work without colliding or repeatedly attempting the same operation.
The report says there was no human operator continuously directing the group from a central console. Instead, each robot perceived its own surroundings, made local decisions, and communicated with the others to negotiate task allocation.
Smaller and more precise operations could be assigned to the desktop robots, while the wheeled humanoid robots could handle movement, transport, and larger assembly steps.
Multi-agent coordination is common in simulation and research papers. Running heterogeneous robots together for many hours in a crowded exhibition environment is harder because communication delays, perception errors, physical obstructions, and accumulated positioning drift can affect the whole process.
The final difficulty was duration.
A robot can look impressive when it performs a carefully selected action for a few minutes. A 15-hour task asks a different question: can the system maintain accuracy and recover from small errors as the number of actions grows?
Every block introduced another opportunity for misalignment. A small error early in the build could affect later sections, so the system needed a closed loop:
The combined test of precision, coordination, and duration was intended to show whether embodied AI could produce useful working hours rather than a short, isolated demo.
The multi-robot challenge drew a crowd at the WAIC 2026 exhibition.
The hardware used in the challenge included the Dexmal Apex, which the company describes as an embodied-native general-purpose robot. Dexmal says Apex integrates its hardware design with an embodied foundation model to support long-duration operation, fine manipulation, broader task generalization, and safe interaction around people.
The robots were the visible part of the demonstration. The underlying intelligence came from Dexmal’s latest embodied foundation model, DM0.5.
Dexmal introduced DM0.5 at its Action 2026 developer event earlier in July. The model is the successor to DM0 and is designed for open-world robot control rather than a fixed collection of scripted tasks.
DM0.5 is designed to generalize across open-world robotic tasks.
The official DM0.5 technical blog says the model uses a Vision-Language-Action architecture with:
The original report describes the model as being trained on approximately 150,000 hours of multi-source data, including real-robot manipulation, first-person human activity, navigation, reconstructed environments, and general vision-language data.
Compared with DM0, the report says the data volume increased by 400% while the parameter count roughly doubled.
The more important change, however, was not simply scale. Dexmal added several mechanisms intended to help the model understand task progress, reason about the physical environment, and learn actions without copying the timing quirks of individual human operators.
Official DM0.5 architecture showing the VLM backbone, context layer, reasoning tasks, and Action Expert.
The original article focuses on three architectural additions. Each one addresses a different weakness in long-horizon embodied tasks.
Traditional VLA models often make decisions from the current camera frame and the current robot state. That can work for short actions, but it becomes fragile when information needed later in the task is no longer visible.
DM0.5 adds a Context Abstraction Layer that combines the current observation with selected visual information from recent history. According to Dexmal, the model can incorporate up to 60 seconds of task context.
During training, the length and availability of the history are varied. If past information is missing or unhelpful, the model is trained to fall back to the current observation rather than failing completely.
Dexmal explains this with a simple experiment: pick up a cup, wipe the area underneath it, and then return the cup to its original location.
Once the cup is lifted, its starting position disappears from the current frame. A model that only sees the present may place it back in an arbitrary location. A model with historical context can remember where the cup was before it moved.
For the Great Wall task, the same principle matters on a much larger scale. The robots need to understand what has already been assembled, how the local structure has changed, and what part of the process should happen next.
DM0.5 also adds 11 autoregressive reasoning tasks to robot-data training.
The model is trained to answer constrained questions using the current image, robot state, instruction, and historical context. These questions cover areas such as:
This changes the training signal from action prediction alone into a combination of:
In practical terms, the robot is encouraged to represent why an action is appropriate, not only which motor command follows a particular image.
That is especially relevant to an 81,920-block task. The system cannot treat each placement as an isolated event. It must connect individual movements to the progress of the larger construction.
Most robot demonstration data is collected through human teleoperation. Different people perform the same task at different speeds, with different pauses and movement styles.
If a model is forced to align every predicted action with a fixed timestamp from the recorded demonstration, it may learn the operator’s rhythm instead of the underlying task.
DM0.5’s Trajectory Alignment Layer aligns actions by task progress rather than exact time.
Each predicted action is matched with an appropriate point in the ground-truth trajectory. The matching must move forward monotonically, so later predicted actions cannot be paired with earlier parts of the demonstration. Dynamic programming is then used to find a low-loss alignment across the full sequence.
This approach is intended to emphasize important physical events such as:
The operator’s personal speed becomes less important, while the meaningful structure of the task becomes clearer.
Dexmal says this improves generalization across different demonstrators and execution speeds.
The Great Wall demonstration was an open-ended public test. Dexmal has also evaluated DM0.5 on standardized robot benchmarks.
On RoboChallenge Table30 V2, the company reports:
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.
Table30 V2 contains 30 real-robot manipulation tasks across four robot embodiments. The benchmark includes long-horizon memory, multi-step execution, object localization, precise pick-and-place, tool use, and bimanual coordination.
The dataset and evaluation framework are publicly documented through RoboChallenge and Hugging Face, while the Table30 V2 methodology has also been published through a CVPR 2026 workshop paper.
The results are relevant to the Great Wall challenge because both settings test whether a model can coordinate perception, task progress, fine movement, and physical execution rather than only recognizing objects or generating text instructions.
A strong base model alone was not enough for a 15-hour public demonstration.
The original report says Dexmal also used DFOL 2.0 during post-training. The framework incorporates the DW0.5 world model, allowing parts of the task to be practiced and refined in a virtual or model-generated environment before full deployment on physical hardware.
The basic idea is to let the system rehearse expensive actions internally.
Traditional robot development often relies on repeated physical trials:
This process is slow and costly. Real robots wear down, physical workspaces must be reset, and every failed attempt consumes engineering time.
A world model can move part of that experimentation into simulation or learned prediction. The robot policy can test possible actions, estimate likely outcomes, and improve task execution before using scarce real-machine time.
For a large construction task, this reduces the amount of trial and error that must happen in front of the audience. It also gives the team a way to rehearse sequences, study likely failure points, and reserve the real hardware for final validation.
The source does not provide a public DFOL 2.0 implementation or a detailed breakdown of how much of the Great Wall task was simulated. The description should therefore be understood as Dexmal’s explanation of its post-training process rather than an independently reproduced result.
Why use more than 80,000 building blocks?
The blocks create a compact test of several capabilities that matter in flexible manufacturing:
Traditional industrial robots are highly effective when a task is repetitive, standardized, and tightly controlled. They are less flexible when the position of an object changes, materials behave differently, or the correct action depends on visual judgment.
The original report argues that a remaining portion of manufacturing work still depends on people because the operations are fine-grained, non-standard, flexible, or perception-heavy. Examples include:
The exact claim that this category represents “the final 20%” of manufacturing should be treated as a framing used in the source rather than a universal industry statistic.
Still, the connection is clear. The Great Wall was a display object, but the capabilities under examination—fine manipulation, sustained operation, and multi-robot coordination—are directly relevant to flexible production lines.
Dexmal was founded in March 2025. The original report says its core team includes former Megvii researchers and engineers, and that the company has received several rounds of investment.
Its stated mission is to build robots that are intelligent, useful, and trustworthy.
Dexmal describes its technical direction as embodied-native development. Instead of adapting a model designed mainly for internet text or images, the company trains models around the needs of robots operating in the physical world.
DM0 was its first embodied-native foundation model. DM0.5 extends that line with longer memory, stronger reasoning, more training data, multi-embodiment support, and improved open-world generalization.
The company has also developed infrastructure around the model.
Dexbotic is an open-source Vision-Language-Action toolbox based on PyTorch. It provides a unified development environment for training and evaluating several VLA policies.
The framework includes:
Dexbotic is publicly available through its official documentation and GitHub repository.
RoboChallenge is a large-scale real-robot evaluation platform co-developed with participation from Hugging Face.
It is designed to make robot-policy testing more reproducible by allowing models to be evaluated on shared physical tasks and hardware configurations.
The platform’s Table30 and Table30 V2 benchmarks cover multiple robot embodiments and practical manipulation tasks. Public datasets, papers, and evaluation resources are available through RoboChallenge and Hugging Face.
Dexmal’s commercial application portfolio also includes Atomix, a flexible warehouse automation business.
The original report says Atomix connects storage, transportation, and picking workflows and has worked with large retail, beverage, and food-service customers. These customer and market-ranking claims come from the report and were not independently audited for this publication draft.
Together, DM0.5, Dexbotic, RoboChallenge, and Atomix represent a full-stack strategy:
The Great Wall installation was the most visible public demonstration of that stack at WAIC 2026.
A public robot challenge does not prove that general-purpose embodied AI is solved.
It does, however, provide a useful test of several abilities that are often evaluated separately:
The most important part of the demonstration was not the wall itself. It was the attempt to keep six physical machines working together long enough for small imperfections and coordination problems to become unavoidable.
Dexmal’s official DM0.5 materials end with the idea that the real world should eventually become the robot’s best teacher. In that vision, every attempt, success, failure, and correction becomes new information that helps the system improve.
At WAIC 2026, the robots were assembling miniature bricks. The longer-term goal is to bring the same closed-loop learning and execution process into manufacturing, logistics, service work, and other environments where tasks cannot be fully scripted in advance.
Six robots participated: four desktop robots and two wheeled humanoid robots. They worked in the same physical area and handled different parts of the construction process.
The model used 81,920 miniature building blocks. The structure was approximately 3.5 meters long, 1.5 meters wide, and 1.1 meters high at its tallest point.
The planned continuous working period was 15 hours. The challenge was designed to test whether the robots could maintain precision and coordination over a long sequence of physical actions.
The original report says the robots were not continuously remotely controlled and did not rely on a completely fixed action script. They used perception, local decision-making, communication, and error correction, although the source does not include a full independent technical audit of the operating setup.
DM0.5 is Dexmal’s open-world Vision-Language-Action foundation model for robot control. It uses a Gemma 3 4B vision-language backbone, a 680M Action Expert, historical context modeling, embodied reasoning tasks, and trajectory alignment.
Dexmal has published the OpenDM repository under the Apache 2.0 license and released DM0.5 model resources through Hugging Face. The repository includes model weights, training and inference scripts, dataset examples, and evaluation workflows.
Table30 V2 is a real-robot benchmark containing 30 manipulation tasks across four robot embodiments. It evaluates abilities including long-horizon memory, sequential execution, fine manipulation, tool interaction, and bimanual coordination.
No. The demonstration shows progress in precision, long-duration operation, and coordination, but it does not establish that all flexible manufacturing tasks are ready for automation. Real production environments require additional safety, reliability, integration, cost, and maintenance validation.
At WAIC 2026, six Dexmal robots worked on a 15-hour challenge to assemble a Great Wall model from 81,920 miniature blocks. The task combined sub-millimeter placement, multi-robot coordination, long-horizon operation, and closed-loop error correction.
The demonstration was powered by DM0.5, Dexmal’s open-world embodied foundation model. Its key additions include up to 60 seconds of historical context, embodied reasoning tasks, and trajectory alignment that reduces the influence of differences in human teleoperation speed.
Dexmal also used world-model-based post-training and connected the demonstration to its broader stack of OpenDM, Dexbotic, RoboChallenge, and warehouse automation products. Several performance and market-position claims remain company-reported rather than independently verified.
The Great Wall was the display, but the real test was whether embodied AI could remain precise, coordinated, and useful for an entire working shift.
Start from one sentence and have a complete website in minutes.
