[Paper Review] L1-Adaptive MPPI Architecture for Robust and Agile Control of Multirotors

Jintasit Pravitra, Kasey A. Ackerman+3 more|arXiv (Cornell University)|Mar 31, 2020

Advanced Control Systems Optimization22 references21 citations

TL;DR

This paper proposes an L1-adaptive MPPI architecture that combines Model Predictive Path Integral Control (MPPI) for real-time nonlinear trajectory planning with L1 adaptive control to enhance robustness against model uncertainties. The L1 controller compensates for mismatches between nominal and true dynamics, enabling stable and agile multirotor flight in simulation despite disturbances like mass changes, inertia shifts, and thrust degradation, with significant improvements in success rate and lap time reduction under extreme conditions.

ABSTRACT

This paper presents a multirotor control architecture, where Model Predictive Path Integral Control (MPPI) and L1 adaptive control are combined to achieve both fast model predictive trajectory planning and robust trajectory tracking. MPPI provides a framework to solve nonlinear MPC with complex cost functions in real-time. However, it often lacks robustness, especially when the simulated dynamics are different from the true dynamics. We show that the L1 adaptive controller robustifies the architecture, allowing the overall system to behave similar to the nominal system simulated with MPPI. The architecture is validated in a simulated multirotor racing environment.

Motivation & Objective

• Address the challenge of robust trajectory tracking in agile multirotor flight under model uncertainties.
• Overcome the inherent fragility of MPPI when simulated dynamics differ from real dynamics.
• Integrate L1 adaptive control as an augmentation to MPPI to ensure robustness without sacrificing planning performance.
• Enable reliable high-speed multirotor racing in complex environments with imperfect dynamics models.
• Demonstrate the effectiveness of the architecture under extreme disturbances such as mass and inertia changes, thrust degradation, and persistent disturbances.

Proposed method

• Use MPPI as a nonlinear model predictive controller to generate optimal trajectories and feedforward controls in real time, leveraging GPU-accelerated sampling.
• Implement a baseline controller using geometric control and quaternion-based PD to correct attitude deviations and improve tracking.
• Introduce an L1 adaptive controller that uses the same nominal dynamics as MPPI to predict and compensate for model uncertainties.
• Design the L1 controller with a state predictor matching MPPI’s nominal dynamics to ensure consistent compensation for mismatched dynamics.
• Apply the L1 controller as an augmentation to the MPPI-generated control sequence, enabling real-time adaptation to unmodeled disturbances.
• Tune the L1 controller with a fast-adapting, robust, and predictable transient response using a 4×4 transfer function and state-space design.

Experimental results

Research questions

• RQ1 Can L1 adaptive control effectively enhance the robustness of MPPI-based trajectory planning under model uncertainties in multirotor systems?
• RQ2 How does the L1-adaptive MPPI architecture perform under extreme disturbances such as 50% mass increase or 100% moment of inertia growth?
• RQ3 Does the L1 augmentation improve mission success and lap time in high-speed multirotor racing compared to MPPI alone?
• RQ4 To what extent does the L1 controller mitigate the impact of thrust degradation and persistent disturbances like nose-up pitching moments?
• RQ5 Can the L1-adaptive MPPI architecture maintain performance when the nominal dynamics do not match the true dynamics, especially in real-time replanning scenarios?

Key findings

• The L1-adaptive MPPI architecture reduced average lap time across all test cases compared to MPPI alone, with the most significant improvement under extreme disturbances.
• In cases with 50% mass increase and 0.1 Nm nose-up pitching moment, the L1-augmented system achieved 100% success rate, while the MPPI-only system failed in all 15 runs.
• For thrust reduction by 40%, the L1-augmented system succeeded in all 15 runs, whereas the MPPI-only system failed entirely, demonstrating critical robustness.
• The L1 controller enabled successful flight through the racing course under all five disturbance scenarios, including those with large inertia and mass changes.
• The architecture maintained stable trajectory tracking and rapid replanning (up to 50 Hz) even when the true dynamics deviated significantly from the nominal model.
• The combination of MPPI’s fast, complex-cost trajectory planning and L1’s real-time adaptation resulted in a control system that behaves closely to the nominal simulation, despite model-plant mismatch.