Shanghai (Gasgoo)- On February 6, Chinese robotics startup Zerith announced the completion of its angel financing round, which raised for the company tens of millions of yuan.
The company also said it has secured a multi-million-yuan order. Besides, its robotics products have entered into the mass production phase.
The company has revealed plans to deliver over 500 humanoid robots this year, targeting sectors such as service, education, entertainment, and exhibitions.
Founded in January 2025, Zerith was jointly incubated by Tsinghua University and the Jianghuai Advanced Technology Center. The company has attracted top talents from major tech and industrial giants, including Baidu, ByteDance, iFLYTEK, and Midea, forming a diverse talent pool spanning fields like the internet, artificial intelligence, and smart manufacturing.
According to the company, Zerith employs AI-assisted structural parameter design to achieve superior dynamic performance and reduced energy consumption for its robots. The team also independently developed a physical interaction world model algorithm for motion control, enabling the robots to adapt to various complex operational conditions and navigate blind spots in unfamiliar environments.
Regarding generalized operational models, Zerith has drawn inspiration from autonomous driving technology to create specialized scene-generalized solutions for robotics. By standardizing tasks in specific scenarios and integrating vision-language large models for scene analysis, the company generates process-oriented task descriptions. During execution, robots dynamically select optimal solutions based on task categories, improving efficiency and adaptability.
Zerith has also categorized robot task types in detail. For long-distance mobile operations like delivery and object organization, the company introduced the GP-6D Universal Grasping and Placement Model. This system leverages vision-language models and 3D scene graphs to deliver collision-free placement postures, addressing a critical gap in existing solutions that focus solely on object grasping.
To tackle cascading error propagation in localized sequential tasks, Zerith's team devised a modular action-primitive-based composite strategy generation framework. This framework decomposes complex tasks into combinable action primitives, which can be selected from a combination library for task completion. To ensure task success, the team proposed an IRL (Inverse Reinforcement Learning) training framework, combining the efficiency of imitation learning and the robustness of reinforcement learning through iterative real-world and simulation training for generalization.
To expand data sources and reduce per-unit data costs, Zerith developed a low-cost monocular camera-based visual remapping robot control solution. Capitalizing on affordable RGB (red, green, and blue) cameras to capture operator movements, the system remaps the actions onto the robot in real time (end-to-end latency of less than 150 milliseconds). This solution overcomes hardware cost limitations, reducing costs by more than 80% compared to traditional controllers or VR systems, and by 99% compared to motion capture technologies, offering a new, cost-effective data acquisition method for remote robot operation.
As for safety, Zerith overhauled its stable version based on extensive user feedback. Anchored on the "Three Safety Laws of Robotics", namely safety for humans, the environment, and the robot itself, the company aims to set a new benchmark for secure humanoid robots.
