iREX 2025: From programmed to perceptive
Physical AI will require a lot of real-world data. A growing share of the hardware used to collect that data – especially low-cost sensing and humanoids – is coming from China.
Key takeaways
- The most recent headlines show a surge in both commercial deployment and ambitious scaling of humanoid robots.
- In late December 2025, Chinese battery giant CATL reported that its “Moz” humanoid robots are now operating on mass‑production EV battery lines, achieving a 99 % insertion success rate by using an end‑to‑end vision model that continuously adapts posture to material‑position variations and precisely gauges force to avoid damaging thin wires.
- At the same time, U.S. startup Galbot announced a $300 million financing round that lifted its valuation to $3 billion; the company is rolling out its humanoid assistants in hospitals to help with patient‑room tasks, pharmacy work and way‑finding, highlighting a shift toward service‑sector applications.
- Meanwhile, the venture‑backed startup Foundation disclosed an aggressive production roadmap that aims to ship 40 robots in 2025, 10,000 in 2026 and a total of 50,000 humanoids by the end of 2027, targeting both industrial and military markets.
- In parallel, industry debate intensified after Figure AI’s Figure 02 model was alleged in a lawsuit to be strong enough to fracture a human skull, prompting calls for clearer safety standards around robot strength.
Physical AI will require a lot of real-world data. A growing share of the hardware used to collect that data – especially low-cost sensing and humanoids – is coming from China. PaXini showed its PMEC Hyper Collection System, consisting of a camera and gloves with multidimensional tactile sensors.
AgiBot, one of China’s leading humanoid robot makers, used iREX to announce its entry into the Japanese market and showed its vision-language-action (VLA) model ViLLA. For example, Yaskawa showed the MOTOMAN NEXT-NHC 10DE, an autonomous dual-arm robot that packs items into a box with human-like delicacy.
According to Yaskawa, the robot’s motions were learned by imitation of a human demonstration. Engineers first had a person wear motion-capture markers on their hands and recorded the person carefully packing a box on camera. Using this captured data, the NEXT-NHC 10DE replicated the human’s packing motions. ## 5. iREX 2025 gives a glimpse into the future
One of my favorite exhibits at iREX came from SOLOMON, a machine vision specialist from Taiwan.
It tuned a Unitree G1 humanoid with enhanced, more capable hands and an onboard NVIDIA Jetson AGX computer. Using NVIDIA’s GR00T platform, the robot was trained to see up to 5 m (16.4 ft.), understand orders in natural language, and plan the physical steps to pick from a defined set of objects.
Mentioned in this article