MIT and Singapore Researchers Give Soft Robots a Nervous System That Adapts on the Fly
A new AI controller modeled on biological synapses lets soft robotic arms handle novel tasks and unexpected disturbances without retraining, achieving over 92% shape accuracy under real-world disruptions.
Overview
Researchers at MIT and the National University of Singapore have developed an AI control system that gives soft robotic arms something closer to a biological nervous system: the ability to learn foundational skills offline and then adapt those skills in real time when conditions change. The work, published in Science Advances on January 9, 2026, addresses three persistent barriers that have kept soft robots confined to laboratory settings despite their potential advantages over rigid machines.
What We Know
The research team — led by Zhiqiang Tang, Liying Tian, Wenci Xin, Qianqian Wang, Daniela Rus, and Cecilia Laschi — drew their design from the way biological neurons maintain stable long-term memories while still updating rapidly in response to new experiences. According to MIT News, their framework uses two complementary sets of connections called “structural synapses” and “plastic synapses.” Structural synapses are trained offline on foundational movements, providing a stable base of skills. Plastic synapses update continuously during operation, fine-tuning the robot’s behavior in real time. A built-in stability safeguard keeps motion smooth and controlled throughout the adaptation process.
As described in the published paper, the offline structural synapses “encode task-agnostic features” while the online plastic synapses are “configuration-specific parameters updated by error-gated rules.” The framework integrates contraction theory with meta-learning, providing formal stability guarantees alongside rapid online adaptation.
The system was tested on two distinct hardware platforms: a cable-driven soft arm and a shape-memory-alloy-actuated soft arm. Results reported by Robotics and Automation News show the controller reduced tracking error by 44 to 55 percent under heavy disturbances and maintained more than 92 percent shape accuracy under payload changes, airflow disturbances, and actuator failures. In a test involving continuously changing fan speeds, the system held 93.8 percent accuracy. Stability was maintained even when up to 50 percent of actuators failed.
The paper also documents that the cable-driven arm “successfully handles payloads reaching 58.5% of its own weight,” a load-to-body ratio that exceeds the 13 to 31 percent range reported in previous studies. The researchers note, however, that control frequency remains relatively low — 2.5 Hz for the cable-driven arm and 1 to 3 seconds per update cycle for the shape-memory-alloy system — due to hardware latency and computational constraints.
Three Problems the Work Targets
According to MIT Global, the controller directly addresses three limitations that have blocked wider deployment of soft robots: the inability to transfer learned behavior from one task to another without retraining, a lack of quick adaptation when environmental conditions shift, and difficulty maintaining stability and safety during behavioral adjustments.
Soft robots — systems built from compliant, deformable materials — can safely interact with humans and handle delicate objects in ways that rigid machines cannot. The challenge has been that their flexible bodies are harder to model and control. Small changes in load, humidity, temperature, or wear can shift how a soft arm responds, quickly invalidating a controller trained for fixed conditions.
“Soft robots hold immense potential to take on tasks that conventional machines simply cannot, but true adoption requires control systems that are both highly capable and reliably safe,” said Daniela Rus, MIT CSAIL director, as quoted by TechXplore.
“This new AI control system is one of the first general soft-robot controllers that can achieve all three key aspects needed for soft robots to be used in society and various industries,” said Zhiqiang Tang, now an associate professor at Southeast University in China, according to Robotics and Automation News.
“This work redefines what’s possible in soft robotics. We’ve shifted the paradigm from task-specific tuning and capabilities toward a truly generalizable framework with human-like intelligence,” said Cecilia Laschi, principal investigator at the Singapore-MIT Alliance for Research and Technology’s Mens, Manus and Machina group, as reported by Robotics and Automation News.
Applications and Next Steps
The research group — which spans the Singapore-MIT Alliance for Research and Technology, the National University of Singapore, Nanyang Technological University, and MIT — identifies assistive robotics, rehabilitation devices, wearable medical devices, manufacturing, logistics, and inspection as target application areas, according to Robotics and Automation News.
The low control frequency remains an acknowledged limitation. The team notes that hardware latency and computational constraints will need to be addressed before the approach can be deployed in time-sensitive settings such as surgery or high-speed manufacturing.
What We Don’t Know
The paper has not yet been independently replicated outside the research group. The two tested platforms — a cable-driven arm and a shape-memory-alloy arm — both demonstrated improved performance, but it is not yet clear how the approach scales to more complex multi-arm or mobile soft-robotic systems. Commercialization timelines and industry partnerships have not been announced.