Humanoid Robots Are Learning to Navigate Obstacles Like Humans

With the aid of a novel system to show them how to sidestep an obstacle, go under a barrier or turn sideways to fit through a tight spot, humanoid robots are improving their ability to navigate cluttered areas. 

The system is the result of work done by Galbot, a Chinese robotics firm. The system is called PASSAGE, and with it a humanoid can modulate its whole body in unfamiliar territory.  

While one might program a robot for discrete motions, PASSAGE is built on a combination of motion planning, whole-body control and environmental perception to determine which movements are needed based on the territory that it has to navigate. To train it, researchers put 100 hours of human movement data from 1,500 virtual settings.  

This was done by having participants in motion-capture gear to make their way around simulated courses, showing how a person will alter his or her posture in confined quarters. The outcome is a system that can plan and carry out its movements without the need for separate directives for every kind of impediment. 

For testing purposes, the team has put a Unitree G1 humanoid through its paces; the modified model has LiDAR sensors for laser measurements and its own onboard computer. The researchers say in 50 physical runs over obstacle courses the robot had never seen before. There was no falling and no call for safety intervention.  

Five of the occasions did see the robot come into contact with something though, so 45 were run clear of any interference. It is a distinction worth making as getting from A to B is not the same as doing so in the vicinity of people or delicate equipment without incident. 

In 750 simulated attempts and in 150 new environments, the system managed to make it to destination 98.67% of the time. That figure dropped to 70.3% when a run only counted as a success if the robot didn’t make contact with a single obstacle. 

Such figures are testament to the advances being made and at the same time to the challenge of instilling precision in a robot’s movements in a variable environment. 

It is a capability of some importance for a humanoid that is to be found in a home, hospital or warehouse. One may have to put a foot over something on the floor, lower oneself to a low ceiling or thread a narrow passage undisturbed. PASSAGE is meant to enable these things from what the robot can sense about its immediate surroundings as opposed to a prebuilt map. 

The technology is still in the experimental stage and the results are of a controlled research nature, not commercial. But it does show how data from human movement can make humanoids more adaptable on their feet, which they’ll need if they’re going to operate in spaces built for people.