Beihang, NTU team proposes PhyFilter in npj Robotics to improve robot generalization beyond data scaling

A team from Beihang University, Beijing Aerospace Control Instrument Research Institute, and Nanyang Technological University's MARS group published PhyFilter in Nature npj Robotics. The module treats a neural network's learning error as a low-frequency signal that can be filtered out, using the robot's real-time state feedback and known physical differential structure to correct learned outputs online; it is plug-and-play, model-agnostic, and its filter parameters can be learned automatically. Validated on quadruped robots, drones, aerial manipulators, and acceleration sensing, policies trained only in flat-ground simulation transferred directly to stone, grass, sand, and gravel, and an aerial manipulator achieved centimeter-level grasping under 5 m/s wind disturbance, suggesting physical filtering feedback as an alternative generalization path to data scaling.

+6