Researchers from the Singapore-MIT Alliance for Research and Technology and the National University of Singapore, working with collaborators from the Massachusetts Institute of Technology and the Nanyang Technological University, have developed an artificial intelligence control system designed to improve the adaptability and reliability of soft robotic arms in real-world environments.
The system enables a soft robotic arm to learn a broad set of motions during an initial training phase and then adapt its behaviour in real time to new tasks and changing conditions without requiring retraining. Soft robots, which are constructed from flexible materials rather than rigid joints and motors, are well suited for tasks that require close and safe interaction with people, but their complex and variable shapes have made them difficult to control consistently outside laboratory settings.
According to the researchers, the new approach addresses three long-standing challenges in soft robotics: the ability to generalise learned behaviours across different tasks, resilience to disturbances such as changes in load or partial hardware failure, and the maintenance of stable and safe motion during on-the-fly adaptation. Previous control methods have typically managed only one or two of these requirements at the same time.
The work is described in a peer-reviewed study published in Science Advances. The control framework draws inspiration from biological neural systems and uses two interacting learning components. One component, referred to as structural synapses, is trained offline on basic movements to provide a stable foundation of skills. The second, known as plastic synapses, updates continuously during operation, allowing the robot to adjust to current conditions. A built-in stability metric constrains these updates to prevent unsafe or erratic behaviour.
Testing was carried out on two different soft robotic platforms, including a cable-driven arm and a shape-memory-alloy–actuated arm. Across these platforms, the system demonstrated reduced tracking errors under external disturbances, high accuracy in maintaining desired shapes despite changes in payload or airflow, and stable performance even when a substantial proportion of actuators were disabled.
Zhiqiang Tang, first and co-corresponding author of the study and now an associate professor at Southeast University, said the system was designed to integrate generalisation, rapid adaptation and stability within a single control framework. Daniela Rus, co-lead principal investigator at SMART and director of MIT’s Computer Science and Artificial Intelligence Laboratory, said the approach was intended to make soft robots more capable of operating safely in unpredictable environments. Cecilia Laschi, principal investigator at SMART and director of the Advanced Robotics Centre at NUS, described the work as a shift away from task-specific tuning toward more generalisable control methods.
The researchers see potential applications in healthcare, assistive and rehabilitation robotics, manufacturing and inspection, where soft robots could adjust automatically to changing tasks or user needs. Future work will focus on extending the approach to faster systems and more complex operating conditions, as well as integrating it into autonomous robotic platforms.
Photo credit: National University of Singapore (NUS)
