PhyFilter from Beihang and NTU lets sim-trained quadrupeds walk real terrain, in npj Robotics
jiqizhixin · x · 2026-09-20
Teams from Beihang University (Academician Guo Lei), Beijing Aerospace Control Instrument Institute (Academician Wang Wei), and NTU MARS published PhyFilter in Nature npj Robotics. A quadruped trained only on flat ground in simulation walks on stone paths, grass, sand, and gravel; an aerial manipulator achieves centimeter-level grasping under 5 m/s wind disturbance — conditions never seen in training.
- Key idea: treat neural network 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 learning output online.
- Engineering: a plug-and-play, model-agnostic lightweight module; filter parameters are learned automatically with no manual tuning.
- Significance: instead of collecting more data to cover every terrain, physical structure is used to filter learning errors online, enabling strong sim2real generalization.
More from Embodied
- Open-source dual-arm robot solves tangram puzzles with 90 demos, ~60% success — DanGrover · 2026-11-12
- Robotics has three regional playbooks: China volume, Europe inspection, US pilots — ingliguori · 2026-11-12
- 3,500 Laser-Armed AI Submarine Robots Patrol Norwegian Salmon Farms — philfung · 2026-11-12
- Model Autonomously Designs a Surgical-Robot-Style Robotic Arm CAD After 75 Minutes — _Stocko_ · 2026-11-12
- Dev releases local world model turning any image into interactive video, runs on RTX 3060 — lucidml_lover · 2026-11-12
- RoboRisk benchmark shows foundation models pose high harm rates in robotics; LEASH harness cuts risk — UoLiverpool · 2026-11-12