Neural-Augmented Incremental Nonlinear Dynamic Inversion for Quadrotors with Payload Adaptation

TL;DR

This work tackles residual dynamics in quadrotor control by replacing or augmenting Incremental Nonlinear Dynamic Inversion (INDI) with learning-based residual predictors. It introduces ILNDI (offline-trained residual learning) and NA-INDI (neural-augmented INDI) to smooth predictions and reduce dependence on noisy or unavailable sensors, extending the approach to cable-suspended payloads. Experiments on a Crazyflie 2.1 show that NA-INDI can outperform pure INDI or ILNDI in some cases and that training on smoothed INDI outputs yields further benefits, though payload complexity can limit gains. The study demonstrates meaningful reductions in sensor requirements and noise in residual control signals, with implications for more robust payload-carrying quadrotor operations and a path toward broader adoption of learning-augmented flight controllers.

Abstract

The increasing complexity of multirotor applications has led to the need of more accurate flight controllers that can reliably predict all forces acting on the robot. Traditional flight controllers model a large part of the forces but do not take so called residual forces into account. A reason for this is that accurately computing the residual forces can be computationally expensive. Incremental Nonlinear Dynamic Inversion (INDI) is a method that computes the difference between different sensor measurements in order to estimate these residual forces. The main issue with INDI is it's reliance on special sensor measurements which can be very noisy. Recent work has also shown that residual forces can be predicted using learning-based methods. In this work, we demonstrate that a learning algorithm can predict a smoother version of INDI outputs without requiring additional sensor measurements. In addition, we introduce a new method that combines learning based predictions with INDI. We also adapt the two approaches to work on quadrotors carrying a slung-type payload. The results show that using a neural network to predict residual forces can outperform INDI while using the combination of neural network and INDI can yield even better results than each method individually.

Figures (5)

• Figure 1: Hardware used for flight experiments. We rely on a Bitcraze 2.1 with standard motors, uSD-card extension, and custom RPM measurement board that uses IR-LEDs. These LEDs can also be tracked by a motion capture system. The payload is equipped with another active IR LED for tracking.
• Figure 2: The first image (a) shows a sample trajectory of those used to generate training data for the neural network, where random points get sampled from a predefined bounding box, to which the quadrotor flies at a random speed from 1 to 8 m/s. The last three images (b), (c) and (d) show the trajectories used for testing the performance of the different controllers.
• Figure 3: The figure shows the outputs of two different MLPs on the residual forces in the x-axis and the original INDI predictions on a Figure8 flight path. Predictions of the MLP when trained on (a) the raw INDI outputs and (b) INDI outputs with spline fitting pre-processing. While the standard MLP has some de-noising properties, it outputs smoother predictions when trained on less noisy data.
• Figure 4: Error comparison between flights with no payload. For Helix, the standard geometric controller (Lee) was unable to fly.
• Figure 5: Error comparison between flights with a payload.