In a world where robots are no longer confined to factories, but instead collaborate with humans, the need for them to possess essential skills in manipulating and interacting with their surroundings has become paramount. Among these skills, pushing has emerged as a key primitive manipulation skill that robots must master.

Imagine a humanoid robot assigned to aid the elderly in their homes, entrusted with the task of retrieving medicine from a high shelf. Instead of laboriously picking up every object obstructing its path, the robot can employ gentle pushes to create a clear corridor towards its target. Pushing becomes even more critical for mobile robots faced with larger and heavier obstacles, particularly in extreme environments such as Mars or during rescue missions after calamities like earthquakes or tsunamis, such as the Fukushima Daiichi Nuclear Power Plant. Moreover, dexterous pushing skills find enthusiastic application and admiration in the world of robot soccer.

While humans effortlessly perform manipulation tasks from a young age, adept at applying learned behaviors to objects of various sizes, shapes, and physical properties, robots face a significant challenge in achieving such versatility. The complexity arises from the fact that physical properties, like frictional forces, are often unknown yet crucial in determining the consequences of a push. Adding to the complexity, pushing dynamics are highly non-linear, with tipping points and sensitivity to initial conditions. Extensive research has been conducted on the mechanics of pushing since the 1980s, producing efficient models for controlling, planning, and predicting the outcomes of pushes. However, models that can generalize to novel objects remain scarce, underscoring the demanding nature of the problem.

Researchers at the University of Birmingham have made significant strides towards enabling autonomous robot manipulators to safely interact with everyday objects, assisting humans effectively. Their research endeavors have focused on endowing robots with the capability to predict the consequences of their actions as they engage in pushing scenarios—a skill humans possess intuitively.

Breaking new ground, Jochen Stuben and Claudio Zito from the University of Birmingham present a comprehensive survey of the current state-of-the-art in robot pushing in their groundbreaking paper, “Let’s Push Things Forward: A Survey on Robot Pushing,” published in the esteemed Frontiers in Robotics and AI journal.

The paper offers an all-encompassing and systematic overview of existing works in push manipulation—a much-needed resource. It primarily targets newcomers, such as Ph.D. students, seeking to understand the evolution of the field. The survey meticulously categorizes a vast array of articles into six distinctive categories: purely analytical approaches, hybrid techniques, dynamic analysis, physics engines, data-driven methods, and deep learning strategies. Additionally, it identifies and discusses five emerging trends in open problems: understanding and semantic representation, sensor fusion and feedback, explicit modeling of uncertainty, cooperative robots and multiple contact pushing, and real-world applications. While the main content of the work delves into a qualitative analysis of the presented methods, the authors masterfully distill complex mathematical concepts into concise, geometrical explanations through helpful figures. Each figure is accompanied by accessible yet meticulous explanations of the mathematical content.

Despite significant advancements, there still remain challenges that require better solutions, while new obstacles and requirements continue to surface in the field. To make pushing a key skill for robots in practical applications, research groups worldwide will need to investigate numerous challenges in the future. Although the theoretical foundations of control and motion prediction are well-established, achieving proper industrial applications remains an ongoing endeavor. While robots in warehouses can navigate freely and deliver goods, the ultimate goal of exploiting pushing operations for novel items in unique situations, such as delicately placing a box of assorted produce on an overhead shelf, is yet to be fully realized.