The Shenzhen-based startup X Square Robot conducted an unedited one-hour broadcast on X and YouTube on August 12th. During this stream, a stationary system with two manipulators was demonstrated loading parcels onto a sorting line. The final counter showed 1,816 parcels with a claimed success rate above 98%.
A more significant metric was the figure displayed on the screen before the count began: 1,248 parcels per hour. This figure closely matches the average rate demonstrated by the Figure AI fleet during its two-hundred-hour autonomous marathon in May. Data analysis shows that 249,560 packages over 200 hours yield an average speed of 1,247.8 parcels per hour. Although X Square did not claim that this benchmark was derived through reverse engineering from the Figure control panel, in a field where any public figure is a message, setting one's own bar based on a competitor's sustained average and subsequently surpassing it by 45% is viewed as a countermove.
Figure's May stream, which started as an eight-hour response to a challenge from industrial automation veteran Dr. Scott Walter, extended to 200 hours. During this test, the humanoid Figure 03 performed a task on a conveyor belt: identifying incoming parcels and flipping them so that delivery labels faced correctly. The system operated under the Helix-02 neural network, achieving approximately 2.6 seconds per parcel and processing about 230 parcels in the first ten minutes of the broadcast without human intervention.
The X Square demonstration covers sorting and flipping, although it does not entirely match Figure's task. X Square's own presentation defines this task as parcel induction: extracting packages from disordered piles left after truck unloading and feeding them one by one onto a conveyor for scanning. Furthermore, the system recognizes and removes irregularly shaped parcels instead of placing them on the conveyor. Observers noted that X Square's task represents a narrower version—grasping and flipping parcels to orient barcodes upwards, as well as directing certain types of parcels into alternative routing channels.
A counterargument is that the stationary, custom-designed cell competes not with humanoids, but with the existing industrial robotics market, including Kuka, Fanuc, Dematic, and numerous Chinese integrators who have been selling six-axis manipulators for logistics for years. In the company's view, X Square's advantage lies not in the hardware, but in the model.
The company insists that at the center of the demonstration is X Square's fundamental embodied AI model, WALL-B. The main thesis is that this model replaces hard-coded rules for each parcel type with a trained policy that evaluates each item and its environment, decides how to handle it, and adapts as the pile changes.
The question of whether a robot for parcel sorting needs two legs and ten fingers remains open. X Square Robot placed a two-armed robot in front of a conveyor belt, equipped with standard industrial grippers. This robot worked with real logistics parcels that varied in size, weight, material, and orientation. During one hour of live broadcast, it performed 1816 effective sorting operations, which is 45% higher than the Figure 03's rate of 1248 parcels per hour.
Both companies engage in similar activities, but their hardware philosophies differ significantly. Figure 03 uses a full-scale humanoid body and five-fingered hands. X Square Robot, however, has abandoned legs and extra fingers, retaining only two manipulators and grippers. As a result, the solution costs approximately 70% less, while maintaining a sorting accuracy exceeding 98%. The parcels processed were not uniform: they included cardboard boxes, soft bags, cylindrical shipments, and fresh goods in styrofoam packaging, simulating a real sorting line.
The key element is the model behind the grippers. In April 2026, X Square Robot introduced WALL-B. This model is built on a single unified architecture that simultaneously trains vision, language, action, and prediction of physical changes within one network, instead of piecing together VLA components. The model determines the next step and predicts the behavior of surrounding objects. Simply seeing a gripper close on a soft bag is not enough; the system must anticipate whether the bag will slip through the middle. When moving a box sideways, it is necessary to assess whether it will shift, rotate, or tip over. This prediction is then used to adjust actions: either by lifting with two arms or by using inertia and the work surface to flip items.
This approach, similar to the idea of DeepSeek, is applied to the hardware: the model takes on the complexity, allowing the body to remain simple and inexpensive. The grippers sacrifice degrees of freedom, and WALL-B compensates for them with smarter strategies dependent on the parcel's state. Hardware still matters: gripper opening, friction, arm load, and force accuracy determine the robot's capabilities. However, now the same standard equipment can handle arbitrary scenes. Figure maintains a fully humanoid form for mobility, tool use, and fine manipulation, whereas X Square Robot supports a minimalist body for the conveyor station, asserting that warehouses should not pay for walking, balance, and ten fingers.
The same model is applicable to both home and industry. After releasing the first comprehensive base model, WALL-A, at the end of 2024, X Square Robot began deploying robots in nursing homes and regular residential homes. In 2026, the company partnered with 58 Daojia to provide on-demand intelligent cleaning services, becoming the first company to implement embodied AI in ordinary homes at scale. WALL-B was tested in these fragmented environments before moving to warehouse settings. Economic factors are what the industry calls the 'killer line': efficiency, cost, stability, and reproduction speed. As models begin to understand weight, friction, and inertia, robots can use fewer expensive bodies, and embodied intelligence approaches overcoming this barrier.
Beijing-based embodied-AI startup X Square Robot has made HOST open-source. This framework enables a humanoid robot to observe a brief demonstration video and successfully replicate the task on its initial try. Using just a 29-second clip, HOST achieved a 62 percent task-success rate, significantly surpassing the industry standard, which typically demands around 50 expert demonstrations and four hours of fine-tuning to reach only 38 percent success.
The prevailing method in embodied-AI involves collecting teleoperation data. In this process, an engineer uses a VR headset to manually guide a robot arm through a task frame by frame, repeating the sequence numerous times before running supervised fine-tuning overnight. The expense associated with this method is substantial; a single data-collection robot can cost hundreds of thousands of dollars, and the marginal cost for one demonstration can exceed one hundred dollars.
While data factories in India and part-time workers in Chinese county towns offer the most affordable source of motion data, the overall cost structure remains difficult to scale.
HOST fundamentally reverses this established procedure. Rather than instructing the robot to mimic human joint movements, HOST prompts the robot to conceptualize the final state and subsequently derive the necessary motion. The framework incorporates two core components: a vision module, which processes the human demonstration to render the scene as the robot perceives it, and an action module, which calculates the joint trajectory required to produce that specific rendering.
Furthermore, a dynamic time-warping alignment module is integrated to map the frames from the human video onto the robot's execution steps, reducing alignment errors by approximately an order of magnitude. This methodology draws inspiration from cognitive science theories of simulation, which posit that humans acquire new motor skills by mentally rehearsing the intended result prior to performing the action.
The Chinese team adapted this cognitive logic for robotic learning, replacing traditional offline fine-tuning with inference-time in-context learning. In this setup, the robot does not undergo retraining on new data; instead, it treats the provided video as a prompt. X Square Robot was established in late 2023 and has since secured over one billion yuan across eight funding rounds, with investors including Meituan Strategic Investment, Alibaba Cloud, Lenovo Capital, and Delian Capital.
The decision to release HOST as open-source represents a strategic move aimed at gaining developer attention as a pathway toward enterprise pilot programs. Although the 62 percent figure serves as a research benchmark and not a ready-to-deploy consumer robot, it effectively redefines the discussion regarding the future architecture of embodied-AI training pipelines once inference-time learning becomes the standard practice.
