Industrial humanoid robots hitting safety wall? Study reveals gap

A new study from Tsinghua University, led by Caiwu Ding and collaborators, exposes a critical barrier to deploying industrial humanoid robots: **they cannot be safely shut down without falling**. The paper, published on arXiv (2608.02809), argues that traditional functional safety standards assume a 'fail-passive' state—where machines safely de-energize upon failure—but legged robots cannot achieve this without toppling. Using a Unitree G1 EDU in a pick-and-place setup, the team mapped out the 'fail-passive gap': classical safety chains (e.g., Siemens S7-1500 PLCs with Stop Category 0) rely on instantaneous power removal, which is impossible for a biped balancing on one leg.

The researchers propose a humanoid-specific safety framework, analyzing active safe states (e.g., single-support stops) and residual risks like fall-as-hazard. They validate their approach in a 3m x 1.5m workspace, but concede that current architectures—even with software-defined automation (SDA) controllers—cannot certify safety beyond the robot’s compute interface. The core issue remains unsolved: **balancing policies and safety runtime cannot coexist on unrated hardware**, leaving humanoid robots in a regulatory deadlock.