New Focus at WAIC: From Single Demos to Continuous Long-Horizon Tasks

At the World Artificial Intelligence Conference (WAIC) held in Shanghai on July 17, 2026, the competitive focus in the embodied AI field is shifting from single, isolated action demonstrations to more complex long-horizon task execution. QX Robotics demonstrated this key transition at its booth, which simulated a home living room, using its Moz1 robot.

When given the ambiguous natural language command “tidy up the living room,” Moz1 did not follow a pre-set script but instead demonstrated autonomous planning and execution. It began by performing a global scan of the environment, then proceeded to place a soda can from the table into the refrigerator and move a dirty bowl to the dishwasher. During this process, when a paper ball was unexpectedly thrown onto the table, Moz1 perceived the environmental change in real-time, paused its original plan to prioritize dealing with the new piece of trash, and then seamlessly resumed its unfinished task of putting away a toy.

The entire process involved multiple sub-tasks across different areas, testing the robot’s ability to remember long-term goals, dynamically re-prioritize tasks, and adapt to interruptions. This marks an evolution for embodied AI systems from solving the ‘how’ of single actions to understanding the ‘what’ and ‘why’ of complex tasks, effectively tackling the industry’s “goldfish memory” problem—losing context or forgetting the ultimate goal in a long chain of tasks.

Spirit v1.6: The Technical Architecture Behind Long-Horizon Tasks

The core technology enabling these capabilities is QX Robotics’ newly released full-stack embodied AI system, Spirit v1.6. The system employs an innovative integrated architecture of a VLA (Vision-Language-Action model) and a world model, designed to allow the robot not only to understand commands and perceive the world but also to predict the consequences of its actions.

Integrated VLA and World Model: The VLA provides an end-to-end connection between human language commands, the robot’s visual perception, and executable actions. The integration of a world model endows the system with a form of “physical intuition,” enabling it to internally simulate and evaluate environmental states and action outcomes. For instance, before attempting to grasp an object, the system predicts the probability of success; after moving, it assesses whether the environment has changed as expected. This comparison between prediction and reality forms the basis for execution monitoring and dynamic adjustment.
Abstract Action Token Prediction: To improve the efficiency and accuracy of long-sequence task planning, Spirit v1.6 introduces the prediction of abstract action tokens during its pre-training phase. Compared to directly generating low-level, lengthy action sequences, abstract action tokens represent higher-level behavioral intents (e.g., “go to the fridge,” “open the door”). This hierarchical approach significantly reduces the computational complexity of long-horizon planning and strengthens the linkage between perception, decision-making, and execution, allowing the robot to plan and adjust tasks at a more macroscopic level.

Through this architecture, Spirit v1.6 establishes a closed loop of “state prediction -> action generation -> execution feedback,” enabling the robot to continuously maintain task status and handle exceptions, rather than blindly playing back a fixed sequence of actions.

Data-Driven Approach and Scenario Validation: From Industrial Lines to the Real World

Technological maturity relies on validation and iteration with real-world data. QX Robotics has built a “data pyramid” system, driving model evolution through multi-layered data sources. The base layer consists of internet video data for building general visual and semantic understanding; the middle layer is human operational data collected via proprietary 7th-generation wearable devices and teleoperation equipment; and the top layer is robot operational data from real-world deployments, especially valuable failure cases.

QX Robotics plans to accumulate one million hours of real-world interaction data by the end of 2026. Currently, its industrial robot, Moz1, is performing complex, non-standard tasks such as high-voltage test plug insertion on CATL’s power battery PACK production lines, achieving a stable success rate of over 99%. Collaborations with companies like JD.com, Bosch, and Schaeffler continuously provide its models with valuable data from retail, industrial, and other scenarios, creating an iterative loop: “real scenarios expose problems -> operational processes accumulate data -> data drives model upgrades -> upgraded models enter more scenarios.”

“Industry First, Commercial Services Second”: Product Roadmap and the Debut of Moz2

Based on its assessment of the technology’s deployment timeline, QX Robotics has established a strategic roadmap of “industry-first, then commercial services, and finally home applications.” Industrial scenarios, with their clear tasks and defined value, along with stringent requirements for stability and precision, serve as the optimal testing ground for refining core technologies.

The Moz2 robot, debuting at this year’s WAIC, marks the next step in this strategy. Moz2 is positioned for commercial service scenarios such as supermarkets, hotels, and office buildings. Its design emphasizes human-robot interaction and environmental adaptability, featuring a cream-apricot colored flexible body, 32 degrees of freedom, an omni-directional wheeled chassis with a folding leg structure for complex terrain, and a 7-DOF biomimetic arm for fine manipulation. Currently, Moz2 is primarily used for exploring human-robot interaction modalities and hardware adaptation in commercial settings, with plans to gradually migrate the long-horizon task capabilities already validated on Moz1.

From the industrially-focused Moz1 to the public space-oriented Moz2, QX Robotics is progressively extending the technological capabilities proven in demanding environments to more open, interactive scenarios, paving a clear path from technical validation to commercial deployment for embodied AI.