Pasadena, California – A humanoid robot dodges a ball thrown directly toward it in a demonstration of technology being developed by Caltech researchers to make robots safer around people and moving objects.

The experiment is part of work by researchers in Aaron Ames’s AMBER Lab, where engineers are developing mathematical safety systems designed to guarantee that humanoid robots can operate safely in the real world. The technology uses control barrier functions to define safe actions and adjust a robot’s movements when hazards are detected.

In one set of experiments, a humanoid approaches a person before stopping and reversing course when it gets too close. Another test shows the robot sidestepping to avoid a person walking toward it, while a dog interacts with the robot and effectively herds it out of the way.

Researchers have also combined the safety system with reinforcement learning and LIDAR sensors, allowing humanoids to detect objects in their surroundings and respond to obstacles, including moving ones.

Ames says modern humanoids can perform increasingly complex movements, but safe operation in unpredictable environments remains a major challenge. He argues that mathematical guarantees are needed to ensure robots do not simply learn how to perform tasks, but can also be prevented from taking unsafe actions.

The AMBER Lab has tested control barrier functions on multiple robotic systems, including drones, quadrupeds and an F-16 fighter jet. Ames says the goal is to allow humanoid robots to eventually perform useful tasks around people while providing a mathematically verifiable layer of safety.