UBTECH Shows Real Factory Work at WRC; DeepSoma Unveils Whole-Brain Simulation

At the UBTech booth, nearly ten industrial humanoid robots, including Cruzr S2 and Cruzr Y1, worked continuously from 9 a.m. to 5 p.m. on tasks such as auto parts loading and unloading, palletizing, and sorting. Without human intervention, the robots performed recognition, grasping, moving, positioning, and retry loops autonomously. The average sorting pace reached nearly 1,100 items per hour, and positioning accuracy was under 1 millimeter. The company said these were not rehearsed demos but exact copies of customer production lines, including challenging operations like handling large workpieces over one meter long with dual-robot coordination.

UBTech's embodied brain is a three-layer system: the Thinker foundation model for understanding, the Thinker-WM world model for predicting consequences, and the Thinker-VLA for execution. According to Jiao Jichao, UBTech's vice president and head of the Embodied Intelligence and Humanoid Robot Research Institute, customers first look at hardware stability, domain knowledge, and whether the solution can actually run, rather than benchmark scores. He said the team has migrated algorithms from x86 to low-power ARM platforms so that the entire software stack can run on a single embedded board. After edge optimization, the VLA model's inference efficiency improved by 176% and storage usage fell by 60%.

Data is another key focus. UBTech said real robot data accounts for 60-70% of its training data, with first-person (Ego) data about 20%, and simulation data the remaining 10%. The company can generate thousands of hours of raw data per day, with about 1.5 to 2 hours of valid data after cleaning every 8 hours. In 2025, UBTech reported revenue exceeding 2 billion yuan, and sold 1,079 full-size humanoid robots, over 80% of which were used in industrial settings such as automobile manufacturing, logistics, and 3C electronics. The company is building a positive loop: deployment in high-value sites generates rich data, which improves models and enables broader deployment.

UBTech also showcased Walker C1, a commercial service humanoid, and U1, a home consumer robot with an emotion large model. Walker C1 was developed from the company's existing industrial robot technology base, taking about 4-5 months from definition to prototype, compared with 6-9 months for a fully new development.

Separately, Zhijue Space Intelligence unveiled DeepSoma, which it describes as a platform to Build Worlds, Run Brains, Embody Intelligence. The company noted that earlier this year, Eon Systems demonstrated a digital fruit fly based on the connectome of an adult fruit fly, involving about 140,000 neurons and 50 million synaptic connections. However, Eon used simplified leaky integrate-and-fire (LIF) neurons and only a small number of manually mapped sensory inputs. DeepSoma aims to go further by modeling biophysical details of individual neurons, including dendrites, membrane potentials, ion channels, and synapses, and by reconstructing real-world scenes into continuously updated, computable 4D digital worlds. The platform also provides a cross-body interface, allowing the same brain model to be embodied in digital animals, biological experiment systems, robotic arms, or humanoid robots.

Zhijue Space Intelligence compares DeepSoma to PyTorch, a common framework for deep learning, but extended to physical environments, brain structures, and agents. The company believes that if this route works, the next leap in physical AI may come not only from larger models and more data, but also from a more ancient and efficient computational prototype — life itself.