Abstract
Where Are We Now?
Since the advent of the first modern robotic devices, such as Unimate, in the 1950s, there has been a consistent advancement in the development of more sophisticated and capable factory machines. This progress has been driven by technological innovations in hardware and computational power. Throughout this period, conventional engineering practices have been applied to the design, construction, and control of these machines, with well-defined constraints and rigorous testing. As a result, numerous applications have emerged, ranging from automotive assembly lines to industrial manufacturing. These robots now operate remarkably well within specified environments and can perform complex tasks with remarkable speed, precision, and efficiency.
However, this progress has primarily been gradual and incremental, as it has successfully met the demands of automation without necessitating any revolutionary changes or relying on untested methodologies. But we are now confronted with the “big problem” of making machines more autonomous, safe, and adaptable, and this continues to pose a barrier to the deployment of robots in everyday life. This is especially true in the field of mobile robotics, where machines must confront changing terrain, unpredictable obstacles, and intricate task planning.
These challenges are exacerbated in untethered machines that have limited power availability or computational speed and bandwidth. Nevertheless, there have been some notable successes. In a recent study, researchers developed an impressive controller that integrates information from different modalities, much like animals do with sensory feedback. 1 Initially implemented in simulation, a privileged learning protocol was utilized to control a quadruped robot. The system was trained to adjust its reliance on sensory feedback from various sources, effectively balancing external inputs (such as Light Detection and Ranging, or depth cameras) and proprioceptive information.
As a result, the walking robot was able to navigate natural terrains, including uphill and downhill pathways and vegetation, successfully and efficiently. However, these approaches have not yet been applied to machines that need to operate in a wider range of environments (for example, burrowing) or possess more flexible or complex architectures.
The Challenge of Flexibility and Deformability
Considering the increasing demand for robots that can safely operate in dynamic and unstructured environments, many researchers have begun exploring more compliant structures that can interact with their surroundings rather than avoiding external contacts. This approach has revealed the inherent limitations of traditional control systems when applied to deformable robots.
Traditional controllers heavily rely on accurate mathematical models of a robot's dynamics, which become increasingly difficult to establish for highly deformable robots. These machines often interact with changing and unpredictable environments where exact modeling is nearly impossible. In addition, the mechanical performance of the robot itself is highly nonlinear. Soft materials rarely have a single stiffness but instead exhibit pseudoelasticity or even viscoelasticity and anisotropy, changing their properties based on the rate and direction of loading. Another issue is that soft robots are designed to adapt, bend, and conform to their surroundings.
Traditional control systems typically depend on linear models and rigid assumptions, lacking the adaptability required to effectively navigate these constantly changing situations. Given that deformable robots are expected to perform in scenarios where traditional machines are incapable or impractical, it is essential that their control systems can be generalized to new situations rather than specific tasks and predefined scenarios.
The Rise of Learning-Based Systems
Despite these challenges, the advantages of compliant mechanisms have prompted a multitude of ideas and approaches from adjacent fields. This includes the utilization of new materials and the incorporation of biological concepts for robot control, such as optimization through genetic algorithms and biomechanical influences in mechanical design. As a result, researchers and engineers are increasingly turning to learning-based systems to tackle the complexities of flexible and deformable robots. A substantial body of work now exists, employing supervised and unsupervised learning algorithms to train robots in specific environments.
Moreover, advancements in adaptability have been achieved through reinforcement learning, where robots learn by trial and error, adapting their behaviors to optimize a given objective instead of relying solely on fixed mathematical models. However, implementing this approach on physical robots can be burdensome, and simulation is limited by the complexity and fidelity required for modeling deformable machines.
Reinforcement learning algorithms can become overwhelmed when faced with numerous decisions and struggle to generalize to different situations. This is where artificial intelligence (AI) driven by deep learning and neural networks has shown remarkable effectiveness in modeling the nonlinear relationships that govern the behavior of flexible robots. These methods empower robots to comprehend their own dynamics and the characteristics of their environment, enabling real-time decision making. This progress has greatly benefitted from advancements in sensor fusion, which allows robots to seamlessly integrate data from various sensors, such as cameras, LiDAR, and tactile sensors.
The Future of Robots in Natural and Man-Made Environments
This combination of traditional robot engineering and advancements in computer science will inevitably result in a broader range of robot deployment. We can expect to see robots performing complex tasks in both natural and man-made environments. For instance, deformable robots, which draw inspiration from the adaptability and resilience of biological organisms, will mimic the locomotion and adaptability of animals. This will enable them to assist in surveying and safeguarding inaccessible parts of the planet.
Robots are also becoming vital tools in various exploratory endeavors, whether it be delving into the depths of the ocean, embarking on remote planetary missions, or navigating the harsh terrains of polar regions. AI-driven control systems will allow robots to adapt to unfamiliar circumstances, gather data, and make informed decisions. Unmanned aerial vehicles and underwater drones will be equipped to handle the challenges posed by turbulent environments and adjust to changing conditions during missions.
Flexible robots will work alongside humans in industries such as manufacturing and construction, providing efficiency while maintaining safety standards. In addition, AI-driven control systems are on track to revolutionize health care. These systems can assist in surgeries, provide patient care, and aid in rehabilitation. These robots are capable of adjusting their movements and forces to suit the specific condition of each patient, ensuring safety and precision.
Expect the Unexpected: A Paradigm Shift in Intelligence?
Although predicting how AI-powered robots will change the world is difficult, if not foolish, there are some important aspects that need to be considered beyond economic and social disruptions. If it becomes possible to implement AI learning in real-time (through a combination of cloud-based and edge computing), robots will learn things that we, as humans, cannot comprehend. The human brain is naturally predisposed to and limited by its sensory systems. It has evolved to process a small range of the electromagnetic spectrum, mechanical stimuli (including audition), and chemical signals.
Machines indirectly enhance our understanding of the universe. For instance, radio wave detectors, electron microscopes, force sensors, and lasers expand our natural set of sensory capabilities. Currently, online AI agents are capable of processing text and imagery provided by billions of people, and we can expect these systems to quickly surpass human analytic capability. Unlike human brains, which have limited bandwidth due to relatively slow electrical signals (individual neurons operate <1 kHz), the processing speed of a computer chip is measured in gigahertz. Even with potential inefficiencies in their architecture, computers will always have a significant speed advantage.
However, computing capacity is not the sole parameter to consider when contemplating intelligence. An important concept that has influenced much of the research in alternative robotics is “embodied cognition,” which recognizes that intelligence emerges from the interaction of the central nervous system, sensory feedback, and morphological computation (the interpretation and processing of interactions through mechanical systems).
The major issue we face is that robots will provide sensory input to massive and fast AI systems that will not be limited to human senses. Instead, AI will have direct access to the full range of detectors invented by humans, extending into the far reaches of the electromagnetic spectrum. Data collection will also expand to include sophisticated mechanical and chemical sensing systems on a scale that is unfathomable. This is tremendously exciting and unprecedented in the natural world. However, an aspect of AI-enabled robots that is somewhat overlooked is that they will interact with and explore this “hyper sensory environment.”
We will no longer have control over the sources of information feeding the AI systems. Instead of AI inputs being confined to human-prepared materials, robots connected to AI will have physical agency; they will manipulate objects in the real world and detect things that are imperceptible to us. It is hard to imagine what these systems will begin to learn and understand. Hopefully, we will have developed mechanisms to compel these systems to provide explanations for their actions!