[Paper Review] Motion Primitives for Robotic Flight Control

Baris Eren Perk, Jean-Jacques Slotine|ArXiv.org|Sep 25, 2006

Control and Stability of Dynamical SystemsEngineering36 references20 citations

TL;DR

This paper proposes a motion primitives framework for robotic flight control using dynamic movement primitives (DMPs) enhanced by nonlinear contraction theory to enable stable, adaptive learning of aggressive UAV maneuvers. By imitating human-piloted flight data and combining DMPs through weighted linear combinations and synchronization, the method achieves obstacle-avoidance flight with high tracking accuracy on a Quanser helicopter, demonstrating robustness and reusability of learned primitives across varying initial conditions and goals.

ABSTRACT

We introduce a simple framework for learning aggressive maneuvers in flight control of UAVs. Having inspired from biological environment, dynamic movement primitives are analyzed and extended using nonlinear contraction theory. Accordingly, primitives of an observed movement are stably combined and concatenated. We demonstrate our results experimentally on the Quanser Helicopter, in which we first imitate aggressive maneuvers and then use them as primitives to achieve new maneuvers that can fly over an obstacle.

Motivation & Objective

• To develop a robust framework for learning and composing aggressive flight maneuvers in UAVs using biological inspiration.
• To address the challenge of controlling highly nonlinear, unstable aerial vehicles in complex, dynamic environments.
• To enable stable concatenation and recombination of motion primitives for new maneuvers, such as obstacle avoidance.
• To demonstrate the feasibility of using human-piloted flight data as a basis for autonomous control via imitation learning.
• To extend DMPs using contraction theory for improved stability under time-varying goals and control parameters.

Proposed method

• The DMP framework uses a canonical system to modulate temporal evolution, with a nonlinear forcing function f(x,v,g) composed of normalized Gaussian basis functions to shape trajectories.
• Contraction theory is applied to ensure exponential stability of the DMP system, even under time-varying goals or control inputs.
• Rhythmic DMPs are extended using a phase oscillator and a stable attractor to model periodic flight behaviors such as hovering or turning.
• Two methods are proposed for generating new primitives: (1) two-way synchronization of DMPs via coupling forces, and (2) weighted linear combination of DMPs' force field weights.
• A two-filter dynamical system (Eq. 31–33) is introduced to improve robustness against time-varying parameters like τ(t) and g(t).
• The method uses experimental data from a human-piloted Quanser helicopter to extract and segment aggressive maneuvers into primitives for recombination.

Experimental results

Research questions

• RQ1 Can dynamic movement primitives be stabilized and composed using contraction theory to enable complex, aggressive UAV maneuvers?
• RQ2 How can human-piloted flight data be effectively segmented and recombined into reusable motion primitives for autonomous flight?
• RQ3 Can synchronized DMPs with coupling forces generate stable, novel trajectories not present in the original data?
• RQ4 To what extent can weighted linear combinations of DMPs produce new, stable flight behaviors with varying start and end points?
• RQ5 How does the proposed DMP framework maintain tracking performance under time-varying goals and system nonlinearities?

Key findings

• The Quanser helicopter successfully tracked the desired trajectory with high accuracy, achieving near-perfect tracking of pitch and yaw angles.
• The roll angle exhibited minor deviations due to control coupling, but followed a consistent pattern, indicating robustness in the presence of multi-variable control.
• Oscillations in the final phase of the roll trajectory were observed, which could be mitigated with improved roll control or velocity observers.
• The combination of sine and cosine DMPs via weighted linear combination successfully generated a new, stable primitive with intermediate phase behavior.
• Synchronization of two DMPs via coupling forces led to exponential convergence, confirming the theoretical stability guarantees of the approach.
• The method enabled the helicopter to perform a complex obstacle-avoidance maneuver not explicitly in the training data, demonstrating generalization and reusability of primitives.