Pioneering research in tip-over stability of autonomous wheeled robots

The study of tip-over stability in wheeled robots is crucial for advancing robotics, especially in outdoor and unpredictable terrains.

The science

Wheeled mobile robots are already indispensable in agriculture, manufacturing, defence, and planetary exploration. On flat, structured surfaces they perform reliably. On unknown, uneven ground they do not, and the most consequential failure is the simplest to describe: the robot tips over. A tip-over can end a mission and damage the machine, and in terrain that has never been surveyed there is no map to warn the robot it is coming.

Existing ways of measuring stability are the starting point, and this research began by testing how far they go. Three established static measures—the Force-Angle Stability Margin, the Tip-Over Moment and Moment Height Stability—each prove useful, but none captures the full range of dynamic effects and terrain interactions that a four-wheeled robot encounters on rough ground. Two further problems had been left open: estimating stability accurately in an environment the robot knows nothing about, and doing something useful with that estimate before the robot goes over.

The research addressed all three. A new stability measure, the Acceleration Angle Tip-Over Measure, accounts for the dynamic factors specific to four-wheeled robots and, unlike its predecessors, expresses stability in quantitative terms a real-time control system can act on. The measure was then extended past the point of no return: a companion measure describes the control actions available for recovering from a tip-over that is already under way.

Making that measure work in the field required two models running together. The first is a physics-based model that uses interoceptive sensors—the robot's measurements of its own state, rather than of the world around it—to estimate the forces and moments acting on it. The second is a deep learning model that uses the same sensor data to predict the robot's support polygon, the shifting footprint that determines whether it stays upright. Combined, the two give real-time stability estimation with no prior knowledge of the terrain at all.

A further deep learning predictor, trained on past interoceptive sensor data, anticipates instability before it arrives. On top of it sits a control strategy that reduces tip-over risk by adjusting the robot's trajectory and speed—and that falls back on the recovery measure when a tip-over is sudden, or when trajectory adjustment alone cannot prevent it.

The complete framework was implemented on a skid-steering four-wheeled robot and tested in the field. Experimental evaluation confirmed it can operate autonomously in unknown environments while maintaining stability, and integrating the control system into a robot significantly reduced tip-over incidents.

Applications

The findings have broad applications across multiple industries. In agriculture, stable four-wheeled robots can be used for precision farming, navigating fields without the risk of tipping and so increasing efficiency and crop yields.

In forestry, stable four-wheeled robots can be employed for tasks such as tree monitoring and habitat assessment, improving operational efficiency and safety in forest management.

In construction, robots with enhanced stability can operate in rugged environments, performing tasks that would be dangerous for humans.

In hazardous areas such as disaster zones or extraterrestrial surfaces, these robots can gather critical data without the risk of toppling, allowing continuous operation.

Because the approach relies on the robot's own internal measurements rather than a survey of the ground ahead, it can be deployed on terrain no robot has crossed before—which is precisely the terrain where tip-over risk is highest.

Impact

The impact of this research extends beyond its immediate practical applications.

By improving the stability of wheeled robots, we can advance automation in sectors that demand reliability in challenging conditions. This raises productivity and improves safety, reducing the need for human intervention in dangerous settings.

There is also a benefit for machines that are not autonomous at all. The approach potentially could be used on manually operated robots/machines which could be valuable as a safety layer on existing equipment, not only on the fully autonomous platforms of the future.

As the robotics industry grows, these advances contribute to more robust and versatile machines, with the potential to transform industries and open new possibilities in automation and artificial intelligence.