How a New Adaptive Solver Cuts Drone Estimation Error by 60% — While Running 70% Faster
For a quadrotor navigating a dense forest at 15 meters per second, a single state estimation error of just 0.3 meters can mean the difference between a smooth flight and a catastrophic crash into a tree. In high-speed autonomous flight, where sensor noise, model inaccuracies, and computational limits collide, the margin for error is razor-thin. Now, a team of researchers at the University of Illinois Urbana-Champaign has developed a new method that slashes estimation error by up to 60% compared to standard approaches — not by throwing more compute at the problem, but by rethinking how optimization itself evolves over time. Their approach, an $\mathcal{L}_1$-adaptive augmented solver for Moving Horizon Estimation (MHE), doesn’t just solve optimization problems faster; it anticipates how they change, adapts to uncertainty in real time, and delivers more accurate state estimates under the very conditions where traditional filters fail: poor initial guesses, high noise, and limited onboard processing. In simulations of aggressive quadrotor maneuvers, the new method outperformed both the Extended Kalman Filter (EKF) and standard MHE, reducing position estimation error from 0.48 meters to just 0.19 meters — a gain that could unlock safer, faster, and more reliable autonomy in drones, self-driving cars, and robotic surgery.
The Science
At the heart of every autonomous system is a silent, continuous calculation: where am I, and what is happening to me? For a quadrotor, this means estimating its position, velocity, orientation, and angular rates from noisy sensor data — GPS, IMUs, cameras — while contending with wind gusts, motor delays, and imperfect models of its own dynamics. The gold standard for this task has long been the Kalman Filter, and its nonlinear cousin, the Extended Kalman Filter (EKF). These methods are fast and elegant, updating estimates recursively with each new measurement. But they have well-known limitations: they can’t enforce physical constraints (like maximum speed or tilt angle), and their linearization of nonlinear dynamics can lead to inconsistent or divergent estimates.
Moving Horizon Estimation (MHE) offers a powerful alternative. Instead of updating a single state estimate, MHE looks backward over a finite window of time — say, the last 20 timesteps — and solves an optimization problem to find the most likely sequence of states and disturbances that could have produced the observed measurements. This sliding-window approach allows MHE to incorporate constraints, handle non-Gaussian noise, and correct for past errors, often yielding higher accuracy than the EKF.
But this accuracy comes at a cost: computation. Traditional MHE treats each time step as a separate, static optimization problem. At every tick of the clock, it discards the previous solution and starts fresh, using iterative solvers like OSQP to converge to the new optimum. On a desktop computer, this is manageable. On a 500-gram drone with a low-power processor, it’s a bottleneck.
The key insight in this paper is that MHE is not a sequence of independent problems — it’s a time-varying one. The optimization problem at time $t+1$ is only slightly different from the one at time $t$: one new measurement has arrived, one old one has dropped off, and the system has evolved by a single timestep. A smart solver shouldn’t start from scratch; it should track the moving optimum, predicting how it shifts and correcting its guess based on new data.
The authors formalize this idea by casting linear smoothing MHE as a dense Quadratic Program (QP) and solving it with a continuous-time Newton’s method — a baseline time-varying solver. But they go further. They recognize that prediction is never perfect: sensor noise, unmodeled dynamics, and linearization errors introduce uncertainty into how the optimal solution evolves. To compensate, they augment the solver with the $\mathcal{L}_1$ Adaptive Optimizer ($\mathcal{L}_1$-AO), a control-theoretic framework originally developed for robust adaptation in uncertain systems.
The $\mathcal{L}_1$-AO acts like a real-time error corrector. It runs in parallel with the Newton solver, estimating the mismatch between the predicted and actual gradient of the objective function. This mismatch — the “prediction inaccuracy” — is fed into an adaptive law that updates a correction term, which is then low-pass filtered to ensure stability. The result is a solver that doesn’t just follow the optimum; it learns the structure of its own prediction errors and compensates for them on the fly.
What They Found
The researchers tested their method in simulation on a quadrotor executing aggressive maneuvers, including rapid accelerations, sharp turns, and altitude changes. They compared three estimators: the Extended Kalman Filter (EKF), standard MHE solved with OSQP (a state-of-the-art iterative QP solver), and their proposed $\mathcal{L}_1$-AO-augmented MHE.
The results were striking. Under nominal conditions, the $\mathcal{L}_1$-AO MHE reduced the root-mean-square (RMS) position estimation error by 42% compared to the EKF and by 28% compared to standard MHE. But the real advantage emerged under challenging conditions — precisely where autonomy systems are most likely to fail.
When the initial state estimate was deliberately set 2 meters off the true position (a common scenario in GPS-denied environments), the EKF took over 8 seconds to converge, during which the quadrotor drifted dangerously off course. Standard MHE converged faster, in about 4 seconds, but still exhibited oscillations. The $\mathcal{L}_1$-AO MHE, in contrast, converged in just 1.8 seconds with minimal overshoot, thanks to its ability to rapidly adapt to the large initial error.
Position Estimation Error Comparison
RMS position estimation error (in meters) for three estimators during aggressive quadrotor flight.
| Label | Value |
|---|---|
| EKF | 0.48 |
| Standard MHE | 0.35 |
| L1-AO MHE | 0.19 |
Figure 1: RMS position estimation error over time during aggressive quadrotor flight. The $\mathcal{L}_1$-AO-augmented MHE (red) converges faster and achieves lower steady-state error than both the EKF (blue) and standard MHE (green), especially after a large initial error at $t=5$ s.
In another test, they increased sensor noise by 50%, simulating degraded GPS or a dirty camera lens. The EKF’s error spiked by 65%, and standard MHE’s by 40%. The $\mathcal{L}_1$-AO MHE’s error increased by only 18%, demonstrating its robustness to uncertainty.
Computationally, the new method was not only more accurate but also more efficient. While OSQP required an average of 12 iterations per timestep to converge, the $\mathcal{L}_1$-AO solver achieved comparable or better accuracy with just one continuous-time update per timestep — effectively a single iteration. This translates to a 70% reduction in computational load, a critical advantage for real-time embedded systems.
Computation Time per Timestep
Average computation time per timestep (in milliseconds) for each estimator.
| Label | Value |
|---|---|
| EKF | 0.8 |
| Standard MHE | 3.2 |
| L1-AO MHE | 1 |
Figure 2: Computation time per timestep (in milliseconds) for each estimator. The $\mathcal{L}_1$-AO MHE (red) runs significantly faster than standard MHE with OSQP (green), approaching the speed of the EKF (blue) while delivering far superior accuracy.
The authors also visualized the internal workings of the $\mathcal{L}_1$-AO, showing how its adaptive correction term (
) rapidly responds to sudden changes in the optimization landscape, such as when the quadrotor enters a sharp turn or encounters a wind gust. The low-pass filter ensures that this adaptation is smooth and stable, preventing high-frequency noise from destabilizing the solver.
Why This Changes Things
At first glance, this work is about optimization algorithms and state estimation — niche topics in control theory. But its implications ripple far beyond the lab. Accurate, robust, and efficient state estimation is the foundation of all autonomous systems. Without it, drones crash, self-driving cars misjudge distances, and robotic arms miss their targets.
Today, most real-world autonomous systems rely on the EKF or its variants because they are fast and simple. But as we push autonomy into more complex, dynamic, and safety-critical environments — urban air mobility, search-and-rescue in collapsed buildings, precision agriculture in windy fields — the limitations of these filters become unacceptable. MHE offers a path to higher performance, but its computational cost has kept it confined to ground vehicles and large drones with powerful onboard computers.
This paper cracks that trade-off. By making MHE not just accurate but also efficient, it brings high-performance estimation within reach of small, agile platforms. A 70% reduction in computation time could allow a delivery drone to run more sophisticated perception algorithms, a surgical robot to update its position at kHz rates, or a Mars rover to navigate rugged terrain with greater confidence.
Moreover, the $\mathcal{L}_1$-AO’s ability to adapt to uncertainty without prior knowledge of the disturbance model is a game-changer. In traditional adaptive control, you often need to assume a specific form for the uncertainty — a slowly varying parameter, a bounded disturbance. The $\mathcal{L}_1$-AO makes no such assumptions. It treats prediction error as a black box and adapts to it in real time, making the system robust to any source of inaccuracy: sensor drift, aerodynamic turbulence, even partial actuator failure.
This is particularly valuable in real-world deployment, where systems face conditions they were never trained or tested on. Consider a drone flying in a wildfire, where smoke degrades camera performance and heat distorts IMU readings. Or an autonomous vehicle encountering black ice, where tire models break down. In such scenarios, the ability to adapt without a detailed model of the disturbance could mean the difference between a graceful recovery and a fatal error.
The method also aligns with a broader shift in autonomy: from static, model-based algorithms to dynamic, learning-enabled ones. While this isn’t machine learning in the deep neural network sense, it shares the same philosophy: embrace uncertainty, learn from data, and adapt in real time. It’s a step toward systems that don’t just follow a script but respond to the world.
What’s Next
The results are promising, but they come from simulation. The next critical step is real-world validation on physical hardware. Quadrotors are ideal testbeds: they’re fast, unstable, and sensitive to estimation errors, making them a stringent benchmark. The authors note that the $\mathcal{L}_1$-AO’s continuous-time formulation may require careful discretization for digital implementation, and the choice of the low-pass filter cutoff frequency $\omega_c$ will need tuning for different platforms.
Another open question is scalability. The current formulation assumes a linear time-varying approximation of the dynamics, updated at each timestep. For highly nonlinear systems — like a flapping-wing drone or a humanoid robot — this approximation may break down, requiring more frequent relinearization or even a nonlinear MHE formulation. Extending the $\mathcal{L}_1$-AO to such cases is an active area of research.
There’s also the question of integration. State estimation doesn’t exist in isolation; it feeds into control, planning, and perception. How does the improved accuracy of $\mathcal{L}_1$-AO MHE translate into better trajectory tracking or obstacle avoidance? Future work could close the loop, showing end-to-end performance gains in autonomous navigation tasks.
Finally, the $\mathcal{L}_1$-AO framework is general. While applied here to MHE, it could enhance other time-varying optimization problems: real-time trajectory generation, adaptive control, even online machine learning. The core idea — augment a baseline solver with adaptive compensation for prediction error — is a powerful paradigm that could redefine how we think about real-time computation in uncertain environments.
In a world increasingly shaped by autonomous systems, the quiet work of estimation — of knowing where you are and what is happening — is more important than ever. This paper doesn’t just offer a better algorithm; it offers a better way to think about adaptation in real time. It’s a reminder that sometimes, the fastest way forward isn’t to compute harder, but to predict smarter — and to learn from every mistake, before it even happens.