Gasgoo Munich- "When a 100-plus-pound humanoid robot suddenly topples over on the factory floor, it's enough to make a safety engineer break into a cold sweat." That was the reality check delivered by Xu Xiaoshun, head of intelligent equipment technology at SAIC-GM Powertrain Technology, on July 3. Speaking at the 2026 Embodied Intelligence Industry Scenario Integration Conference hosted by Gasgoo, Xu highlighted the practical hurdles facing humanoid robots in industrial settings.
A decade ago, Xu worked on a humanoid project at GM North America designed for space stations. It was expensive and never meant for mass production. "It wasn't mature back then," he said. "Now, the domestic supply chain is fully developed, and the conditions are in place." Today, the market is crowded with domestic brands like Unitree, Zhiyuan, UBTECH, Galaxy General, and Magic Atom. While Tesla continues to push forward overseas, it is actually China's new energy vehicle makers that are flocking to join the R&D race.
Xu Xiaoshun, Head of Intelligent Equipment Technology at SAIC-GM Powertrain Technology Co., Ltd.
Xu envisions a future where the ratio of industrial robots, collaborative robots, humanoids, and human operators reaches 1:1:1:1. For now, however, he sees wheeled humanoids as the smarter transitional choice. "Under the new national standards, both bipedal and wheeled designs are classified as humanoid robots," Xu noted. "From a factory perspective, wheeled models offer higher efficiency and provide a buffer for both the physical environment and worker psychology."
SAIC-GM's roadmap runs through virtual benches, simulation training, physical testing, POC validation, on-site verification, and multi-unit iterations — all aimed at achieving human-machine collaboration. The company has successfully deployed a humanoid robot for loading new energy battery cells on a mass production line, marking the first such application at SAIC Motor. Operating on a single shift, the robot loads over 1,000 cells daily, hitting 80% of human capacity while significantly improving ergonomics.
Yet Xu admits that stability and MTBF (mean time between failures) remain the biggest pain points for bipedal models. Battery swapping requires professional handling, and lost teleoperation signals can risk motor overload. That gives wheeled robots a distinct advantage at this stage. Looking ahead, large embodied AI models could tackle longer, more complex tasks — but only if hardware stability improves, costs fall, and certification systems mature. "We need lower costs and higher stability," Xu said. "Combined with large models, that's how embodied intelligence will truly reach the factory floor."
