SMART and NUS weaponize AI to give soft robots human-like reflexes

The system reduced tracking error by 44% to 55% under heavy disturbances.

The Singapore-MIT Alliance for Research and Technology (SMART) and the National University of Singapore (NUS) have developed an AI control system that enables soft robots to adapt to new tasks and disturbances in real time without retraining, according to a press release.

The system reduced tracking error by 44% to 55% under heavy disturbances and maintained over 92% shape accuracy when payloads changed or actuators failed, with stable performance even when up to half of the actuators were disabled, the study said.

The controller combines structural synapses trained offline on foundational movements with plastic synapses that update during operation, allowing a single model to generalise across tasks whilst maintaining stability during adaptation.

The approach was validated on both cable-driven and shape-memory-alloy–actuated soft robotic arms, which demonstrated accurate trajectory tracking, object placement and whole-body shape regulation under airflow disturbances and mechanical failures.

The results show the system maintained controlled performance during real-time adaptation across different soft-arm platforms, including scenarios involving actuator failure.