At the WAIC 2026 exhibition in Shanghai, Efort Qizhi demonstrated a full-size humanoid robot navigating uneven, gravel-strewn surfaces modeled after lunar terrain. The robot approached an irregular boulder, bent down, lifted it securely with both hands and carried it several meters to a designated area.
Human Teaching Through Multi-Modal Data Collection
Visitors not only watched pre-programmed demonstrations but also tried on the HALO human skill collection suit developed by Efort Qizhi. The suit simultaneously records visual, force, tactile and language signals while operators perform tasks, allowing the system to convert human actions into training data. This approach differs from traditional robot programming that relies on fixed code written by engineers and instead focuses on distilling real human operations into general representations.
The collected data includes both successful and failed attempts to broaden generalization, combined with decades of industrial process documentation and standard operating procedures. Through a human-to-robot action transfer model, skills are mapped across humanoid robots, composite robots and industrial arms, solving the problem of data silos between different hardware configurations.
HumanGPT World Model and Complete Toolchain
Efort Qizhi developed HumanGPT, a specialized world model for embodied intelligence that differs from general large language models such as Doubao or Qwen. HumanGPT fuses data from the HALO suit with multi-modal sensor inputs to construct a virtual dynamic space where robots can plan and execute actions. The model works together with the Dayan data platform, Modou IDE and Openmind OS to form a complete technology stack from data acquisition to on-site deployment.
Chairman Dr. You Wei described the skill iceberg concept, noting that visible simple skills such as grasping and sorting represent only the tip, while the submerged portion includes complex industrial requirements such as multi-robot interaction, tightly coupled process parameters and real-time synchronization. Examples include autonomous generation of spray angle, flow rate, welding current and grinding spindle speed when performing coating, welding or polishing tasks.
Cross-Scenario Skill Transfer Demonstrated Live
During the exhibition, an Efort Qizhi humanoid robot received an unprepared task to move and install a large photovoltaic panel. Without dedicated programming or tuning for this scenario, the robot used HumanGPT and Openmind OS to identify obstacles, plan a path across uneven ground, carry the panel while its view was partially obstructed, and finally use a dexterous hand to operate a standard electric screwdriver to secure the panel with screws.
The company deliberately chose challenging, non-standard scenarios to prove genuine capability rather than relying on simplified setups. The same system has already been deployed in over 4000 factories, accumulating more than 5000 hours of continuous operation on customer production lines.
By providing a standardized universal technology base, Efort Qizhi aims to reduce repeated low-level development across the industry and enable traditional automation integrators and end users to deploy embodied intelligence solutions without deep expertise in AI algorithms or robot control.
