Stairs are one of the most difficult environments for robots to navigate safely.
Even robots specifically designed to climb stairs can lose balance, and when they do, the results can be serious.
A heavy robot tumbling down a staircase can damage itself, damage property, and potentially injure nearby people.
Now, researchers at the Singapore University of Technology and Design (SUTD) have developed a new safety system that helps robots catch themselves during a fall.
Instead of focusing only on preventing accidents, the system teaches robots how to recover when a fall becomes unavoidable.
The research, published in Results in Engineering, uses artificial intelligence to control a robotic arm that acts like a protective brace during a fall.
According to the researchers, fall prevention alone is not enough. A robot may be moving safely when an unexpected event occurs, such as a person accidentally bumping into it on a staircase. In situations like these, a fall may be impossible to avoid.
To address this problem, the team equipped a commercial stair-climbing robot with a three-jointed arm mounted at the rear of the machine. The arm was designed to quickly move into position and help stabilize the robot if it begins to tip over.
The researchers first studied how robots typically fall on stairs. They identified five common types of falls, including backward falls, sideways falls, and falls involving rotation. After analyzing these scenarios, they determined that a three-jointed arm was the simplest design capable of helping the robot recover from all five situations.
Rather than programming specific movements by hand, the team used reinforcement learning, a type of artificial intelligence that learns through trial and error.
Inside computer simulations, the robot was repeatedly pushed in different directions. The AI controller had to decide how to move the arm in fractions of a second. Successful recoveries were rewarded, while crashes and unstable movements were penalized. Over thousands of practice sessions, the AI gradually learned more effective recovery strategies.
The results were encouraging. The AI-controlled system successfully stopped falls and returned the robot to a stable position nearly 70% of the time. A traditional hand-coded control system succeeded less than 40% of the time and often made the situation worse by moving the arm ineffectively.
When the AI system successfully recovered from a fall, it usually stabilized the robot in just over four seconds.
The researchers also tested whether the system could work on different robots and staircases. A controller trained on one robot design was able to function on larger and smaller versions without additional training. In some cases, performance even improved, reaching a success rate of 87%.
This suggests the system is learning a general recovery strategy rather than memorizing a specific staircase or robot design.
Although the technology is not yet ready for real-world deployment, the researchers see it as an important step toward safer autonomous robots. Future work will include testing on physical robots and combining the AI system with other safety measures such as brakes and fall-prevention systems.
The ultimate goal is to create stair-climbing robots that people can trust to operate safely in hospitals, office buildings, warehouses, and other environments where stairs remain a major challenge.
