| May 29, 2026 |
A reinforcement-learning-based safety system teaches a stair-traversing service robot to brace itself mid-fall, addressing one of the biggest barriers to deploying autonomous robots on staircases.
|
|
(Nanowerk News) Staircases are among the most challenging terrains a mobile robot can face. A multi-year field study found that robots designed for stair traversal fail at least 35 times more often on stairs than on level ground. The consequences can be significant. A robot that loses balance on a step accumulates momentum as it tumbles, threatening severe damage to itself, the building, and anyone in its path.
|
|
Fall prevention measures, such as path planning and balance control, can help the robot avoid noticeable hazards, but they cannot stop a person from accidentally walking into the robot from above. That residual risk is unavoidable, and it has kept operators from deploying heavy autonomous platforms in stairwell environments.
|
|
“This is why fall mitigation matters more than fall prevention alone,” said Professor Mohan Rajesh Elara, who heads the Robotics and Automation Research (ROAR) Laboratory at Singapore University of Technology and Design (SUTD). “Until the industry has a credible answer to that residual risk, operators will keep treating these platforms as a liability rather than a labour-saving tool.”
|
|
In their study (Results in Engineering, “A reinforcement-learning-based fall mitigation system for stair-traversing service robots”), Prof Elara and his team at ROAR developed a fall mitigation system for robots through the use of design, artificial intelligence (AI), and technology. They equipped a commercial-grade tracked robot with a three-jointed arm controlled by a policy trained entirely through reinforcement learning (RL) in simulation.
|
 |
| Overview of the proposed fall mitigation system. (Image: Singapore University of Technology and Design) (click on image to enlarge)
|
|
By surveying how falls happen on stairs, the team identified five fall modes: a straight backwards fall, two pivoting fall variants, and two sideways falls. They then explored the simplest type of articulated structure that could brace against all modes.
|
|
“Three degrees of freedom turn out to be the minimum that can geometrically cover the five fall types, when the structure is mounted at the rear of the robot,” explained Prof Elara. “The mechanism narrows the problem enough that an AI controller can solve it. The AI lets us drive a mechanism that would otherwise be too complex to control by hand.”
|
|
In each simulated episode, a force knocked the robot backwards or sideways, and the controller decided every fraction of a second how to move the arm joints. A proximal policy optimisation algorithm then adjusted the controller’s behaviour based on the outcome by rewarding stable, upright finishes, while penalising flips, falls off the stairs, or unnecessary arm movement.
|
|
Across five trained controllers, the RL system achieved an average success rate of 69.4 percent in arresting the robot’s fall and returning it to a stable position. In contrast, a hand-coded heuristic baseline succeeded only 38.6 percent of the time and often destabilised the robot further by flailing the arm. When the RL controller did catch a fall, it stabilised the robot in an average of 4.25 seconds— well within the 10-second window the team set as their target.
|
|
The most commercially significant result came from robustness testing. The controller, trained on a single robot and staircase geometry, was tested without retraining on platforms that were either 10 percent larger or smaller, and on staircases with altered step dimensions. On a larger robot, the best controller’s success rate climbed to 87 percent. On a smaller, less stable platform, it dropped but still functioned.
|
|
“The controller is not memorising one geometry,” added Prof Elara. “It is learning a recovery strategy that generalises.” This means a single module could be reused across a family of robots sharing the same morphology, without retraining for each variant.
|
|
The researchers acknowledge that an average success rate of 69.4 percent falls short of what safety standards like the IEC 61508 demand for a standalone safety function. For the solution to reach deployment readiness, the team needs to raise the controller’s own performance, add compensating measures such as mechanical brakes and upstream fall-prevention layers, and meet explainability requirements through a surrogate model. A physical validation will first entail a simplified test rig, then integration onto a full platform, followed by testing across real staircases of varying geometries.
|
|
This study is part of a longer programme at SUTD’s ROAR Lab that advances the safe operation of reconfigurable mobile robots. The work is supported by two national programmes in Singapore that are both oriented towards translating research into deployable platforms.
|
|
“Our goal is to make the system one credible layer of defence within a larger safety architecture. If we can demonstrate that it works reliably and is auditable, we move stair-traversing robots closer to being platforms that operators trust rather than fear,” said Prof Elara.
|
