Knowing Their Limits

Robotics Institute Studies How Humanoid Robots Can Safely Stop in Emergencies

The Breakdown

• Malfunctioning humanoids can hurt people, and improved safety measures are essential as these robots become more common in daily life.
• CMU and Siemens partnered to create a method to determine the “safe stoppability” of humanoids.
• They also defined conditions where it’s safe to stop a humanoid using simulation, resulting in Proactive Refinement of Importance-sampled Stoppability Monitor (PRISM).

* * *

Videos of humanoid robots failing — running into barriers, falling down or even accidentally kicking a child (who was not seriously hurt) — rack up hundreds of thousands of views online. 

Changliu Liu, a faculty member in Carnegie Mellon University’s School of Computer Science, and members of her Intelligent Control Lab watch them as well. But they see these failures as dramatic examples of the dangers humanoids pose as they increasingly operate near people and other robots.

“We see that humanoids can hurt people,” said Liu, an associate professor in CMU’s Robotics Institute. 

Improved robot safety measures are essential as humanoids become more common in workplaces, public spaces and even homes.

Liu and her students partnered with colleagues at Siemens to develop a method for assessing the “safe stoppability” of humanoid robots. Researchers use this method to monitor humanoids and alert human operators when a robot may be unable to stop safely. But Liu and Yifan Sun, the robotics Ph.D. student who led the project, envision a day when humanoids could internalize this method and use it to keep themselves out of danger.

Emergency stop (E-stop) mechanisms have been the de facto standard for robot safety, Liu said, and many robots have a big red button mounted on them that is easily accessible. 

For an industrial robot that is mounted in place, pressing the E-stop — the big red button — cuts power and halts all motion. But a humanoid requires power to balance and control its joints, so simply cutting power to it will cause it to collapse, potentially harming itself or anything (or anyone) it falls on.

So rather than cutting power, roboticists have sought to stop humanoids by first directing the robot to assume a neutral position. In the CMU study, that neutral position was the robot standing with its arms hanging down on either side. Once in the neutral position, additional steps, such as attaching the robot to a support harness, could commence and allow operators to safely power down the robot.

“But it’s not always safe to execute this fallback to the neutral position,” Liu said. 

The robot might be surrounded by people or other obstacles, could be operating on a slippery or uneven surface, or may be holding an object. In that case, returning to a neutral position could be risky. 

Context, therefore, is everything. So the researchers set out to define operational conditions in which a humanoid might be safely stoppable. Outside of this safety envelope of actions and environments, the robot might not be able to stop safely. 

“For instance, when the robot approaches a conveyor belt with a human worker nearby, its current state may become dangerous, and an operator may want to stop the robot or move it to a less risky place,” Sun said.

The researchers employed simulations to explore which environments and activities might impede a stop. Simulation enabled them to study rare but critical states and to do so safely without damaging a robot. The result was the Proactive Refinement of Importance-sampled Stoppability Monitor (PRISM). 

Siemens demonstrated PRISM using a commercially available robot that stands a bit more than four feet tall. In the tests, the researchers found that PRISM enabled proactive safety monitoring and could support certification of fail-safe behaviors, a particular interest of Siemens.

Though currently designed as a monitor for a human controller, Sun said PRISM could be used by the robot itself to recognize high-risk situations and to avoid or minimize any time spent outside the safety envelope. This awareness could help identify additional risky situations and lead to improved, safer operation.

Liu recently discussed robot safety on CNN.

In addition to Liu and Sun, the research team included Yiyuan Pan, a master’s student in robotics, and Shangtao Li, a master’s student in mechanical engineering. It also included three members of Siemens’ Foundational Technologies Division: Caiwu Ding, Tao Cui and Lingyun Wang. Siemens and the National Science Foundation supported the study.

The research will be presented at IROS 2026, the International Conference on Intelligent Robots and Systems, Sept. 27-Oct. 1 in Pittsburgh.

For More Information: Aaron Aupperlee | 412-268-9068 | aaupperlee@cmu.edu