L1-MPPI: L1 Adaptive Model Predictive Path Integral for Agile UAV Control

arXiv:2609.38467 · cs.RO · Submitted 2026-09-29 · Read on arXiv

Listen

Transcript

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

Rosa: Today's paper: "L1-MPPI: L1 Adaptive Model Predictive Path Integral for Agile UAV Control".

Dev: L1-MPPI proposes an L1 Adaptive Model Predictive Path Integral framework that cascades L1 adaptive control with MPPI to enhance trajectory tracking for high-speed UAVs under model uncertainties and external disturbances.

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: Moving on to the actual summary of "L1-MPPI: L1 Adaptive Model Predictive Path Integral for Agile UAV Control," the core idea is that they cascade L1 adaptive control with MPPI to boost trajectory tracking capabilities for high-speed UAVs. Dev So, in simpler terms, it's taking an existing path integral method and supercharging it with an L1 adaptive augmentation to handle things we can't perfectly predict beforehand.

Taro: That sounds like they are trying to make the planning part of the system more robust against surprises rather than just relying on a perfect initial guess of the dynamics. What kind of surprises are they specifically targeting?

Rosa: They are targeting model uncertainties, which includes things like an additional payload or a mismatch in how aerodynamic drag is modeled, which is something many current MPPI approaches overlook. This capability means the tracking remains accurate even when those external factors are present.

Dev: So, it's not just about trajectory following anymore; it’s about maintaining that precision even when the physical system deviates from the assumed model due to real-world physics or unknown conditions. That pushes us toward a more resilient control architecture.

Taro: I think the paper emphasizes that they are not just adding a simple correction term; they are building a framework where the L1 observer estimates remaining model uncertainties online and compensates for them in real time, which is what makes it adaptive.

Rosa: That adaptability is powerful because it means the system can adjust its control actions dynamically as conditions change, rather than relying on pre-programmed responses to known errors. It's a continuous adjustment process.

Dev: From my point of view, that online compensation capability is interesting because we have to be careful about how fast the L1 controller reacts; the paper mentions a lowpass filter for the estimated matched uncertainty to yield the adaptive command, which suggests they've put some guardrails on its response speed.

Taro: And I also noticed they model communication delays explicitly within this structure, which ties everything together—the high-level planner, the flight controller, and the estimator are all accounted for in one cohesive loop.

Rosa: It seems like a very holistic approach; by incorporating these low-level details and adaptation into the path integral formulation, they are tackling the tracking problem from both a planning perspective and an estimation perspective simultaneously.

Dev: That simultaneous modeling of dynamics, control structure, and uncertainty estimation is what makes it different from systems that might just treat those as separate modules.

The paper's summary: Rosa: Now let's talk about the specific improvements detailed in "L1-MPPI: L1 Adaptive Model Predictive Path Integral for Agile UAV Control." The authors highlight that their approach improves upon existing MPPI by explicitly modeling the low-level flight controller, motor dynamics, and communication delays. Dev That level of detail in the model seems to be a major step up from what we typically see when we design these kinds of high-speed control loops.

Taro: Modeling those low-level dynamics is critical because it means they can anticipate how the motors will actually respond to the commands calculated by the path integral, which helps in achieving more physically accurate and stable inputs during aggressive maneuvers.

Rosa: Beyond just modeling dynamics, a key improvement is that they use an iterative mixing scheme that reflects how a low-level controller operates when updating the nominal control sequence based on sampled disturbances. This connects the high-level planning back to the underlying control actuation in a more integrated way.

Dev: That iterative mixing scheme sounds like it addresses one of those common issues where you have a decoupled planner and executor; they're ensuring consistency between what's planned and what the motors are actually capable of doing.

Taro: And then, they add the L1 adaptive augmentation which compensates for uncertainties online, meaning if the model drifts or an unknown payload hits, the system actively corrects itself without needing a full re-planning cycle every time. That continuous adaptation is a significant enhancement over purely nominal models.

Rosa: So we're looking at three main improvements: better modeling of low-level hardware, an iterative mixing strategy for control updates, and the addition of real-time online uncertainty compensation via L1 augmentation.

Dev: And when you look at the evaluation, they specifically tested this controller onboard in real-world flight rather than just in simulation only, which is a huge validation point for its robustness under dynamic conditions.

Taro: The paper points out that this framework enables agile flight in the presence of model uncertainties and external disturbances, which is exactly what we need for autonomous systems to function reliably when things get messy outside the controlled lab setting.

The paper's improvements: Rosa: So, wrapping up our discussion on "L1-MPPI: L1 Adaptive Model Predictive Path Integral for Agile UAV Control," the main implication is that this framework offers a way to achieve robust trajectory tracking for high-speed UAVs even when facing real-world uncertainties like unknown payloads or drag mismatches. Dev It really shows how integrating low-level dynamics and adaptive compensation can significantly enhance the performance of MPC-based methods compared to standard implementations.

Taro: I think the paper’s contribution lies in showing that explicitly modeling the low-level controller and motor dynamics, combined with an L1 adaptive augmentation, provides a way to maintain agile flight while actively compensating for unmodeled system variations without needing constant re-planning.

Rosa: And from my perspective as a field roboticist, the fact that they validated this onboard in real-world flight—achieving speeds up to thirteen point five m/s and accelerations up to two point five g with a positional RMSE reduction of fifty-eight point six one percent compared to plain MPPI under an unknown payload—is what really makes this paper stand out for practical use.

Dev: I agree; the performance numbers they show, especially that significant reduction in positional error when an unknown payload was present, demonstrate that this method is very effective in dynamic conditions where robustness is essential.

Taro: For the future work mentioned, it sounds like they are focused on pushing this further by investigating how this structure can be applied to even more complex scenarios than just simple payload changes or drag mismatches.

Rosa: That sounds like a solid direction; I’m really looking forward to seeing how this L1-MPPI framework can be adapted for even more challenging, unstructured flight environments in the future.

Dev: It's definitely a method that pushes the boundaries of what we expect from standard path integral control methods when dealing with high-speed, uncertain dynamics.

Conclusion: Rosa: So, to wrap up our discussion on "L1-MPPI: L1 Adaptive Model Predictive Path Integral for Agile UAV Control," we've seen how this framework improves trajectory tracking by incorporating low-level dynamics and online adaptation. Dev It really shows how integrating those low-level details and adaptive compensation can significantly enhance the performance of MPC-based methods compared to standard implementations.

Dev: It's definitely a method that pushes the boundaries of what we expect from standard path integral control methods when dealing with high-speed, uncertain dynamics.

Luka´s Kotek, Ond ˇ ˇrej Prochazka, Voj ´ ech Von ˇ asek, Martin Saska, Robert P ´ eni ˇ cka ˇ

Czech Technical University in Prague

cs.RO

Submitted: 2026-09-29

Updated: 2026-09-29

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: L1-MPPI proposes an L1 Adaptive Model Predictive Path Integral framework that cascades L1 adaptive control with MPPI to enhance trajectory tracking for high-speed UAVs under model uncertainties and

Key concepts

Model Predictive Path Integral (MPPI): MPPI uses forward simulation to generate candidate control trajectories based on a reference path. It evaluates these paths using a cost function that measures how well the UAV follows the desired trajectory, then selects the best nominal control input from these sampled possibilities.
L1 Adaptive Augmentation: L1 augmentation compensates for model errors and external disturbances by estimating uncertainties in real-time. It uses an observer to estimate these uncertainties and a low-pass filter to generate an adaptive control signal that corrects the system's performance against known or unknown mismatches.
Cascaded Control Architecture: The L1-MPPI framework combines two distinct control methods: MPPI handles high-level trajectory planning, while L1 adaptive control provides robust compensation for model uncertainties. The L1 command is added to the MPPI command to produce the final, enhanced flight control input.
Low-Level Dynamics Modeling: The system includes detailed models for both the UAV's main dynamics and its low-level flight controller, including motor dynamics and communication delays. These models are crucial for accurately simulating how physical components respond to control inputs before the high-level MPPI optimization takes place.

Terminology

Adaptation — 20× in this paper · explained in 1 episode
Model Predictive Path Integral — 18× in this paper
Core — 6× in this paper · explained in 2 episodes
Robustness — 5× in this paper · explained in 5 episodes
Adaptability — 4× in this paper · explained in 1 episode
Model Predictive Path Integral (MPPI) — 4× in this paper · explained in 3 episodes
Ablation — 2× in this paper · explained in 1 episode
Accuracy — 2× in this paper · explained in 1 episode
Consistency — 2× in this paper · explained in 3 episodes
Dynamic Modeling — 2× in this paper · explained in 3 episodes
PID Controller — 2× in this paper · explained in 2 episodes
Ablation study — 1× in this paper · explained in 1 episode
Graphics Processing Unit — 1× in this paper
Model Drift — 1× in this paper · explained in 4 episodes
Model Reference Adaptive Controller — 1× in this paper
Online Adaptation — 1× in this paper · explained in 3 episodes
Root Mean Square Error — 1× in this paper

Summary

L1-MPPI proposes an L1 Adaptive Model Predictive Path Integral framework that cascades L1 adaptive control with MPPI to enhance trajectory tracking for high-speed UAVs under model uncertainties and external disturbances. This method is significant because it improves accuracy even when facing unknown payloads or mismatches in aerodynamic drag, a capability not fully realized by existing MPPI approaches.

The gist

The proposed L1-MPPI controller cascades the L1 augmentation with MPPI instead of the NMPC used in [10], which keeps the sampling-based cost of MPPI while targeting the same agile flight regime. Compared to the L1-augmented MPPI of [14], we additionally model the low-level PID controller, motor dynamics, and communication delays, and we evaluate the controller onboard in real-world flight rather than in simulation only.

UAV Model and Low-Level Dynamics

The control architecture begins with two core models: the dynamic model of the UAV (Section III-A) and the model of its low-level flight controller (Section III-B). The nominal UAV dynamic model describes the state as position, velocity, body rate, and quaternion. The equations governing these dynamics include terms for linear aerodynamic drag approximated by a diagonal matrix D [17]. A key contribution is considering both the UAV model and its flight controller to achieve precise trajectory following using MPPI (Section III-D).

The low-level flight controller dynamics are modeled as:

e = ωc − ω, e˙I = e, τc = kpe + kieI

The normalized allocation matrix G is used to compute individual motor throttle commands from the collective throttle and target torques. The motor dynamics are then modeled using a first-order model with a time constant kmot (Equation 4).

Model Predictive Path Integral (MPPI)

Trajectory following is performed by the MPPI, which computes nominal control inputs such as thrust FMPPI and body-rate ωMPPI based on the reference trajectory Tref. The principle of MPPI control involves generating a set of candidate trajectories (rollouts) using forward simulation of the system dynamics:

u k j = u nom j + δu k j

The cost function C k evaluates how precisely the UAV follows the reference trajectory Tref. The weights wk are then given by:

**/w k = exp − 1 / λ (Ck − ρ) **

Subsequently, the nominal control sequence is updated using a weighted average of the sampled disturbances:

/u nom j:= X K k=1 wk · uk j

L1 Adaptive Augmentation

The L1 adaptive augmentation compensates for model uncertainties and external disturbances (e.g., unknown or changing payload) by providing compensating thrust FL1 and body-rate ωL1. The system dynamics are rewritten to separate nominal and uncertain components:

/v˙ = 1/m R(q)0 0 Ft + σ − DRT(q)v + g, ω˙ = J−1(τ − ω × Jω) + ξ

The system dynamics are separated into nominal and uncertain components as:

/z˙ = f(R(q))+g(R(q))(uL1+σm)+g⊥(R(q))σum, (16)

The L1 observer estimates the state zˆ, and the piecewise-constant adaptation law estimates the uncertainties σˆ = [σˆm, σˆum]T. The L1 control law applies a lowpass filter to the estimated matched uncertainty to yield the adaptive command:

/uL1,k = uL1,k−1e − ωcoTs − σˆm,k(1 − e − ωcoTs)

Integration and Real-World Evaluation

The final control command is computed by adding the L1 command to the MPPI command:

/u d = uMPP I + uL1,k

To account for communication delay, the iteration is propagated forward using the last applied command:

/x0 = x + fRK4(x,ulast, ∆t)

The separation of L1 Controller and MPPI ensures consistency by subtracting the L1 thrust from the measured forces before initializing the MPPI prediction. The results demonstrate that under an additional payload, full L1 augmentation outperforms methods relying solely on online mass estimation in flight. Real-world experiments achieved speeds up to 13.5 m s−1 and accelerations up to 2.5 g with a positional RMSE reduction of 58.61 % compared to plain MPPI when an unknown payload was present.

Improvements for AI systems

Here are specific improvements for AI systems based on the proposed L1-MPPI controller, and what those improved systems can achieve:

Improving UAV Trajectory Tracking under Unknown External Loads: The system can now maintain high-speed, agile flight (up to 13.5 m/s) while accurately tracking complex trajectories despite significant model uncertainties, such as an unknown payload (up to 35% mass increase).
Enhancing Robustness Against Aerodynamic Mismatches: The system maintains superior tracking performance even when aerodynamic drag coefficients are mismatched or uncertain, a scenario where existing methods often fail.
Integrating Low-Level Dynamics for Realistic Control: The improved AI system can explicitly model and compensate for the dynamics of the low-level flight controller and motor dynamics, leading to more physically accurate and stable control inputs during aggressive maneuvers.
Achieving Onboard Adaptive Compensation: The system features an L1 adaptive augmentation that estimates remaining model uncertainties (like unknown disturbances) in real-time, allowing it to compensate for these changes without requiring complex, pre-programmed adaptation laws or relying solely on external mass estimators (which degrade under high drag mismatch).
Mitigating Control System Delays: The architecture explicitly accounts for communication delays between the high-level planner and the low-level actuators, ensuring consistent performance even when the control loop has inherent latency.
Enabling Real-World Agile Flight: The system is validated for real-world deployment (e.g., using Jetson Orin NX) to achieve high accelerations (up to 2.5 g) and demonstrate its adaptability under dynamic, non-nominal conditions that are difficult to capture in pure simulation.
Optimizing Control Allocation Under Saturation: The controller incorporates an iterative motor-mixing strategy at the throttle level, prioritizing collective thrust preservation during actuator saturation events (like motor desaturation), ensuring the UAV maintains essential stability and flight capability.