Actuator-Aware Observer Inputs for Cascade ADRC of Quadrotors Under Rotor Lag, Measurement Delay and Sensor Noise

summary

The gist

In sampled cascade linear active disturbance rejection control (LADRC) of quadrotors under rotor lag, measurement delay and sensor noise…

In short

The research addresses errors in quadrotor control systems caused by rotor lag, measurement delay, and sensor noise within cascade linear active disturbance rejection control (LADRC). It proposes actuator-aware observer inputs: using filtered virtual inputs for altitude/attitude and reconstructed acceleration for horizontal motion. These designs significantly reduce tracking errors compared to traditional PID methods.

Key concepts

Actuator-Aware Observer Inputs
This is a design where the observer receives specific, tailored information based on the system's dynamics. For altitude and attitude, it uses virtual inputs filtered by an assumed rotor model aligned with measurement delays. For horizontal motion, it uses acceleration reconstructed from measured attitude instead of just instantaneous tilt commands.
Sampled Cascade LADRC
This is a control strategy used for quadrotors that combines multiple control loops in a specific sequence. It is designed to reject disturbances by using an observer that estimates the system's state, which then feeds into the controller. The 'sampled' aspect refers to how measurements and commands are taken over time.
Rotor Lag
Rotor lag refers to the delay between a command sent to a motor and the actual resulting change in rotor speed or acceleration. This physical delay introduces inconsistencies between what the observer expects and what is actually measured, creating structural residual errors that this research aims to fix.
Covariance Ceilings
These are mathematical constraints placed on the control variance (how much uncertainty is allowed) and its increment within each control channel. They act as tuning limits to ensure the observer design remains robust and stable, preventing noise or estimation errors from causing runaway control actions.

Terminology used across episodes

This episode discusses

The paper

Actuator-Aware Observer Inputs for Cascade ADRC of Quadrotors Under Rotor Lag, Measurement Delay and Sensor Noise · Read on arXiv

Eastern International University · Industrial University of Ho Chi Minh City

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Actuator-Aware Observer Inputs for Cascade ADRC of Quadrotors Under Rotor Lag, Measurement Delay and Sensor Noise".

Dev: Here's the gist of the paper. In sampled cascade linear active disturbance rejection control (LADRC) of quadrotors under rotor lag, measurement delay and sensor noise, actuator-aware observer inputs are proposed to address structural residual errors caused by input inconsistency The main contributions include two actuator-aware observer-input paths for altitude and attitude, a fixed-gain covariance design with activity ceilings, a paired nonlinear simulation assessment, and a repetition in PX4 software-in-the-loop Observer Input Design Paths The paper introduces two distinct observer-input paths for sampled cascade LADRC: altitude and attitude observers use limited virtual inputs filtered by an assumed rotor model aligned with the measurement delay, while horizontal observers use acceleration reconstructed from measured attitude For altitude and attitude, the controller reconstructs normalized inputs from the limited thrust and torque before rotor allocation, propagates them through a first-order actuator model, and aligns the result with the measurement time For horizontal motion, the observer uses near-hover acceleration reconstructed from measured attitude rather than the instantaneous tilt command The proposed actuator-aware formulation for altitude and attitude propagates the limited virtual command through an internal rotor-lag model, where u k is driven by aˆk−nˆd, αˆ = e−Ts/τm For horizontal channels, the observer is driven by near-hover horizontal-acceleration surrogate reconstructed from the delayed attitude measurements Tuning and Constraint Formulation The tuning employs fixed gains selected on the discretetime model under sensitivity and step-response requirements, alongside stationary covariance ceilings on each channel’s control variance and increment, as well as on the noise-induced inner-loop increment of the full cascade The nominal covariance model includes the closed attitude loop, its observer, actuator lag, and measurement buffers to predict tracking error and noise-induced virtual-control variance Sensitivity requirements constrain the search such that Ms ≤ 1.98, followed by a final audit against Ms ≤ 2 The outer-to-inner noise path becomes an explicit design constraint, as the cascade-level increment σ∆ui,cas is predicted on the full cascade model and agrees with nonlinear hover measurements Performance Assessment and Results The paired set-point trials show that PID-Budget, AABudget and AA-HalfGain reach mean RMSEs of 65.5, 19.2 and 16.9 mm, respectively Under a stricter common cascade ceiling, the AA designs reach 29.9 and 26.4 mm versus 72.3 mm for a PID that uses roughly double their rotor-command increment The AA-Budget-E design reduces the set-point RMSE from 59.1 mm (PID-E) to 44.0 mm with a lower rotor-command increment in every paired set-point and lemniscate run Validation and Robustness Checks PX4 software-in-the-loop runs reproduce these results within 2 mm in two independent campaigns Behind EKF2, a noise model fitted to the estimation errors predicts the roll/pitch and yaw increments within 5% on held-out and out-ofsample runs and the total rotor-command increment within 14% The evidence is simulation-based, with results showing that AA-Budget reduces mean position RMSE by 70.7% relative to PID-Budget while reducing measured rotorcommand increments This paper showed that the input supplied to the extendedstate observer is a design variable in sampled cascade LADRC of quadrotors Delay-aligned actuator-model inputs for altitude and attitude and measured-attitude inputs for horizontal motion remove the nominal structural observer-input residual under matched lag and delay With commanded inputs, feasible designs were found, within the stated finite search, only at the lowest observer/controller ratios, with 9.

Paper summary: Rosa: So, this paper is about this thing called "Actuator-Aware Observer Inputs for Cascade ADRC of Quadrotors Under Rotor Lag, Measurement Delay and Sensor Noise". Basically, they’re looking at how to fix errors when you’re using a quadrotor with sampled cascade linear active disturbance rejection control.

Dev: And the core problem is that the extended state observer usually gets the command right away, but the measurement you actually get is delayed and filtered because of rotor lag.

Rosa: Right, so there's this mismatch between what your controller thinks should be happening and what you're actually seeing from the sensors. This paper proposes a way to fix that by making the observer inputs smarter about how things are delayed.

Dev: They claim they use two different observer input paths: one for altitude and attitude, and one for horizontal motion. The idea is that altitude and attitude observers get inputs filtered by an assumed rotor model that’s timed to match the measurement delay.

Taro: And for those horizontal channels, they aren't just using the tilt command directly; they reconstruct acceleration based on what you *do* measure from the attitude sensor. That seems like a way to compensate for the delay in a specific way.

Rosa: Exactly, it’s about aligning the inputs with that known lag and delay. Then, they tune everything—the fixed gains—based on sensitivity and step-response requirements, and they put these constraints on how much variance or increment each channel can have due to noise or the control loop itself.

Dev: They also look at noise in a specific way: they set stationary covariance ceilings on the control variance and the increment for every single channel, plus this extra constraint on the inner-loop increment of the whole cascade. That’s where you manage how much noise can blow up things down the line.

Taro: I'm interested in those constraints because when you're dealing with real hardware, knowing exactly where that noise boundary is is crucial for graceful recovery when things go wrong.

Rosa: And they show some results in paired set-point trials, where they compare different designs like PID-Budget, AABudget, and AA-HalfGain. They see mean root mean square errors of sixty-five point five millimeters for the altitude error, nineteen point two millimeters for attitude error, and sixteen point nine millimeters for horizontal motion.

Paper summary: Dev: That’s a good baseline number to start with when you’re trying to reduce tracking errors in real-time loops. But they show that under a stricter common cascade ceiling, their designs like AA-HalfGain reach twenty-nine point nine and twenty-six point four millimeters, compared to seventy-two point three millimeters for a PID that uses roughly double the rotor command increment.

Taro: So, it’s not just about getting lower errors; it’s about how much better you can get when you respect those noise constraints on the entire system. What does that imply for a system that's running in real-time?

Rosa: It implies that by designing the inputs to be actuator-aware, you can push the performance much closer to what a well-tuned PID would achieve, even under heavy lag and delay conditions.

Dev: They also have this follow-up with EKF2. They fitted a noise model based on the estimation errors, and this model predicts roll and pitch increments within five percent on held out and out of sample runs, along with the total rotor command increment within fourteen percent.

Taro: Fourteen percent for the total increment sounds like a decent bound for how much you can expect things to drift when you’re pushing the system. That gives us a clearer picture of the uncertainty in those predictions.

Rosa: And they show that by redesigning it specifically for this estimator-fed loop, using estimator-aware increment ceilings, they reduce the set-point error from fifty nine point one millimeters with PID E down to forty four point zero millimeters.

Dev: That reduction is significant because it also results in a lower rotor command increment in every single paired set-point and lemniscate run. They even found a forty five percent reduction before quantization and fifteen percent after output quantization for the mean S six rotor command increment relative to PID E.

Taro: So, the paper seems to be saying that by being aware of the actuator dynamics—the lag, the delay—you can actively manage those increments so they don't just accumulate noise.

Paper summary: Rosa: And they validate this in PXfour software-in-the-loop runs where everything was code generated into a real autopilot. They also show that AA-HalfGain beats PID E only on the lemniscate run, which is a specific stress test.

Dev: The evidence they present is simulation based, and the ceilings they use are just tuning constraints, not some kind of magical fix. But seeing those numbers in PXfour SITL runs within two millimeters in two different campaigns gives it some real weight.

Taro: What this means for someone actually flying a drone is that you have to consider these kinds of input designs when you're building the control stack, especially if your hardware has known lag issues.

Rosa: Right, so the paper "Actuator-Aware Observer Inputs for Cascade ADRC of Quadrotors Under Rotor Lag, Measurement Delay and Sensor Noise" shows that by being specific about how you feed information to your observer based on actuator awareness, you can get tighter control margins.

Dev: It points toward a design where the input isn't just the instantaneous command anymore; it’s an input tailored to account for the known physical delays and noise characteristics of the system.

Taro: And if you look at their limitations, they note that these results are based on simulation and real-world validation in PXfour so you still have to be careful when translating those specific numbers directly to a completely novel hardware setup.

Rosa: That’s fair; it’s simulation-based evidence, and the tuning constraints they used were just nominal tuning parameters. But the idea that AA-Budget reduces mean position error by seventy point seven percent relative to PID-Budget while cutting down rotor command increments is pretty compelling.

Dev: It suggests that this approach isn't just about making things look better on a chart; it’s about fundamentally reducing the uncertainty in the control loop's execution when you have those inherent system delays.

Taro: So, for the real world, it means when you’re designing a real-time loop, you can afford to be more explicit about how your observer handles input timing and noise budgets.

Rosa: And that’s what this paper is about—finding those specific, actionable inputs for the observer so the system doesn't get bogged down by rotor lag and sensor noise.

Conclusion: Rosa: So we're wrapping up this look at "Actuator-Aware Observer Inputs for Cascade ADRC of Quadrotors Under Rotor Lag, Measurement Delay and Sensor Noise." Basically, they figured out a way to fix those annoying residual errors when you've got quadrotors running sampled cascade control.

Dev: Yeah, the authors focused on how the observer gets its inputs—altitude and attitude—instead of just taking whatever command is handed to it instantly.

Taro: They proposed these two different input paths, one for altitude and attitude that filters inputs based on an assumed rotor model that matches the measurement delay.

Rosa: And for horizontal motion, they used acceleration reconstructed from the measured attitude instead of just the tilt command directly. It's about aligning what you see with what you know is happening physically inside the rotor system.

Dev: The tuning part is interesting too, because they set fixed gains based on sensitivity requirements and put limits—called activity ceilings—on how much variance or increment each channel can have due to noise or the control loop itself.

Taro: I liked how they modeled that inner-loop increment for the full cascade, because that tells you exactly where the noise is likely to blow up down the line.

Rosa: The results showed they could lower those tracking errors significantly—down to around twenty-nine point nine millimeters under stricter constraints—compared to a standard PID setup.

Dev: And they validated this in software like PXfour showing that their designs reduce the rotor command increment, which is key for stability when you have those kinds of physical delays.

Taro: It means the system is more robust when it's actually flying outside of a perfect simulation environment.

Rosa: The whole point here is that the input to your extended state observer isn't just a command; it’s a design choice that directly manages how well the quadrotor handles lag and noise.

Dev: It gives us a clearer path for designing controllers in real-time where latency and sensor noise are constant factors.

Taro: Next time we talk about this, we can look at how these observer designs hold up when the world gets really messy or unpredictable.

More episodes

← Home