PhyFilter: physics-informed filter lets sim-trained robots traverse real terrain without extra data
机器之心 · wechat · 2026-09-11
A team from Beihang University and NTU's MARS Lab published PhyFilter in Nature npj Robotics, offering an alternative to data scaling for robot generalization.
The method treats learning residuals as a low-frequency signal that can be filtered: using the robot's real-time state feedback and known physical differential structure, it corrects network outputs online. PhyFilter is plug-and-play, model-agnostic, learns its filter parameters automatically via optimal control and adjoint gradients, and runs at 500Hz on an STM32 microcontroller.
Validated on quadrupeds, drones, aerial manipulators, and acceleration estimation: a policy trained only in flat-ground simulation traversed flagstone, grass, sand, and gravel (up to 80% success vs near-zero baselines); drone tracking error dropped 30.22%; an aerial arm completed centimeter-level pick-and-place under 5m/s wind. Paper, code, and website are open-sourced.
More from Embodied
- Replit Agent in a robot builds and publishes websites autonomously via MCP — amasad · 2026-09-11
- TARS Robotics unveils embodied foundation model AWE: 15+ tasks, one model, zero retraining — heyshrutimishra · 2026-09-11
- Qualcomm's next-gen Hexagon NPU runs 30B MoE models with 32K context on-device — lee_stott · 2026-09-11
- Working with an ESP32 device using Copilot CLI and Astra — DanWahlin · 2026-09-11
- VATIX: open-source driving world model trained on 5,500 hours of real-world footage — abursuc · 2026-09-11
- Reachy Mini robot gains spotlight after joining NVIDIA's portfolio — RachelVT42 · 2026-09-11