Zero Power founder Zhang Zhening, a graduate of Tsinghua AI Lab, demonstrated its core technology, the ZERITH-H1 wheel-arm humanoid robot. Its anthropomorphic upper limbs integrate 28 degrees of freedom and are equipped with hyperspectral visual and tactile sensors (derived from the lab's ICRA-awarded technology). These sensors simultaneously capture data in seven modalities, including vision, force, touch, and joint motion. The accompanying VR control system reduces human-machine latency to 3ms, allowing engineers to control the robot in a virtual "twin-like" manner, enabling four hours of continuous data collection without interruption.
Even more groundbreaking is its "virtual-reality self-reinforcement learning" architecture: through a Real2Sim2Real loop, a high-fidelity simulation environment is first trained with real-world scenario data, which is then fed back into the model using millions of operational data points. Finally, errors are corrected in the real world through self-correcting RL. Field data showed that the Zerith-V0 model maintained a 99.8% success rate after 72 hours of continuous operation in a desktop cleaning task, even in complex scenarios such as mixed oil and water stains and obstructions, far exceeding the industry average of 85%.
Unlike most companies' concept machine demonstrations, Zero Power directly brought vertical scenarios such as "home cleaning" and "industrial component assembly" to the conference. Its data management platform is compatible with mainstream algorithm frameworks such as ACT and Diffusion Policy, allowing customers to import data with one click and quickly fine-tune models. The team has reportedly partnered with three leading home appliance companies and plans to launch a "modular cleaning robot solution" by the end of the year, reducing deployment costs per scenario by 60% compared to traditional solutions.
"We're not making a 'walking PowerPoint,' but rather letting robots 'learn by doing,' just like humans," Zhang Zhening said in an interview. Zero Power's "omnimodal" architecture not only addresses the current data shortage but also provides expanded interfaces for force, touch, semantics, and world models, laying the foundation for the evolution of large-scale embodied models over the next 3-5 years. This "scenario-driven R&D" approach aligns with the Ministry of Industry and Information Technology's strategic positioning of embodied intelligence as "shifting from virtual reasoning to physical manipulation."
As humanoid robots rapidly penetrate a wide range of industries, Zero Power's WRC debut is not only a testament to its technological prowess but also a sign that China's post-2000 entrepreneurial generation is redefining the commercial future of embodied intelligence through a dual approach of "deepening application development and bringing technology to everyone."