Researchers at the Singapore University of Technology and Design (SUTD) have developed an AI-driven safety system that enables stair-climbing robots to brace themselves during a fall. Published in the journal Results in Engineering, the study addresses a critical barrier in autonomous robotics: the high failure rate of machines navigating stairs. Field data indicates that robots fail at least 35 times more often on stairs than on flat ground, posing significant risks of damage to infrastructure and people due to the momentum gained during a tumble. Professor Mohan Rajesh Elara, head of the Robotics and Automation Research (ROAR) Laboratory at SUTD, emphasized that traditional prevention methods like path planning cannot eliminate all risks, such as human interference. Consequently, the industry requires a robust mitigation strategy to make these platforms viable for deployment. The team created a system using a three-jointed arm mounted on a commercial tracked robot. Through simulation-based reinforcement learning, the robot learns to respond to five distinct fall modes: straight backward, two pivoting, and two sideways falls. The researchers determined that three degrees of freedom were the minimum mechanical requirement to geometrically cover these scenarios. The reinforcement learning controller operates by adjusting the arm's joints in real-time fractions of a second after a simulated fall force is applied. Utilizing a proximal policy optimization algorithm, the system is rewarded for stable recovery and penalized for further instability or unnecessary movement. Across five trained controllers, the AI system achieved a 69.4% success rate in arresting falls and returning the robot to a stable position, significantly outperforming a hand-coded heuristic baseline which succeeded only 38.6% of the time. Successful recovery occurred within an average of 4.25 seconds, well within the team's 10-second target window. A key commercial advantage of this approach is its robustness. The controller, trained on a specific robot and staircase geometry, demonstrated the ability to generalize without retraining. When tested on platforms 10% larger or smaller and staircases with altered dimensions, the system maintained high functionality, achieving an 87% success rate on a larger robot. This indicates that the AI is learning a universal recovery strategy rather than memorizing specific conditions, allowing a single module to be adapted across various robot families. Despite these successes, the researchers acknowledge that the 69.4% success rate currently falls short of the stringent IEC 61508 safety standards required for standalone safety functions. Future steps involve improving controller performance, integrating mechanical brakes, and adding upstream prevention layers. The team also plans to address explainability requirements through surrogate models and will proceed with physical validation on test rigs before full-platform integration. This research is part of a broader initiative at SUTD to advance the safe operation of mobile robots. Supported by national programs in Singapore aimed at translating research into deployable platforms, the goal is to create a credible layer of defense. By demonstrating that the system is reliable and auditable, the team aims to shift operator perception of stair-traversing robots from liabilities to trusted labor-saving tools.