Robot learningPhysical AI

PhyFilter uses robot physics to correct learned policies when conditions change

A September npj Robotics paper reports a model-agnostic physics filter tested on quadrupeds, drones, aerial manipulation and acceleration estimation under unseen conditions.

By TechniaHQRobot6 min read
PhyFilter tests across quadruped locomotion and aerial robotics

Introduction

A paper published in npj Robotics on September 4 proposes PhyFilter. The method adds a physics-based correction to the output of an existing learned model when the robot encounters conditions outside its training distribution.

The authors test it across quadruped locomotion, drone flight, aerial manipulation and acceleration estimation. They report real-time execution on an STM32F765 at 500 Hz in the aerial cases.

Source review date recorded for this article September 14, 2026.

The correction sits outside the learned policy

PhyFilter uses known differential structure plus live state feedback to estimate a residual correction. The paper presents it as a plug-in module that can work with pretrained policies without retraining the network.

That architecture supports before-and-after comparisons under the same disturbance.

Tests include conditions excluded from training

The reported experiments include terrain, payload and speed changes for locomotion plus wind and mass uncertainty for aerial systems. In one drone experiment, the paper reports mean absolute tracking error reductions of 30.22% relative to SEER-I and 50.17% relative to the baseline under combined mass and wind uncertainty.

For aerial manipulation, the authors report a maximum manipulation error of 2.5 cm under 5 m/s wind and 0.3 kg mass uncertainty, with average and variance improvements of 23.99% and 55.04% over baseline. These are paper results under the stated tests, not universal guarantees.

Small processors are relevant for deployment

Running at 500 Hz on an STM32F765 shows that the correction equations fit on a resource-constrained controller in the tested setup.

Deployment still needs bounds for bad sensors, incorrect physical parameters and actuator saturation.

A residual dynamics correction can fail when sensors or actuator limits are wrong

PhyFilter is reported as a model-agnostic correction layered over a learned controller. That separation is useful because a pretrained policy need not be retrained for every mass or wind change, but the corrective model depends on the state supplied to it and the physical assumptions used in the residual calculation. If the estimator is delayed or an actuator is already saturated, a mathematically sensible correction may not be executable.

For a physical replication, publish the pretrained policy, the mass and disturbance ranges, sensor sampling method and whether correction computation missed any deadlines. Compare trajectories with and without the filter under identical controlled disturbances, including out-of-range cases. The paper's reported 500 Hz embedded execution measures the tested correction implementation; it is not a universal rate for every robot's control stack.

Conclusion

PhyFilter is a useful example of combining learned policies with explicit robot physics when deployment conditions move outside the training set.

Sources and methodology

Physics filtering favors the generalization of robot learningnpj Robotics · September 4, 2026
Article by @techniahqrobot