From Flies to Robots: Inverted Landing in Small Quadcopters with Dynamic Perching

TL;DR

This work tackles dynamic inverted landing for small quadcopters by teaching a general two-stage policy that converts augmented optical-flow cues into triggering and rotational actions. It combines RL-derived discrete sensory-motor pairs with a continuous mapping in an augmented optical-flow space $\mathbb{OF}_{a}$, using an OC-SVM for triggering $\pi_{Trg}$ and a FFNN for rotation $\pi_{Rot}$, and demonstrates zero-shot sim-to-real transfer aided by domain randomization and system identification. Key contributions include the bulk RL data collection across varied ceiling-approach conditions, the formalization of the triggering boundary and rotation mapping, and an analysis of how landing-gear geometry influences inverted landing robustness. The findings show robust inverted landing behaviors in hardware and reveal practical design trade-offs between leg geometry and approach conditions, offering a path toward autonomous dynamic perching in cluttered environments. $\mathbb{OF}_{a}$, $\tau$, $\vartheta_x$, $D_{ceil}$, $s_{Trg}$, $a_{Rot}$, $\pi_{Trg}$, and $\pi_{Rot}$ are central to the method and its transferability.

Abstract

Inverted landing is a routine behavior among a number of animal fliers. However, mastering this feat poses a considerable challenge for robotic fliers, especially to perform dynamic perching with rapid body rotations (or flips) and landing against gravity. Inverted landing in flies have suggested that optical flow senses are closely linked to the precise triggering and control of body flips that lead to a variety of successful landing behaviors. Building upon this knowledge, we aimed to replicate the flies' landing behaviors in small quadcopters by developing a control policy general to arbitrary ceiling-approach conditions. First, we employed reinforcement learning in simulation to optimize discrete sensory-motor pairs across a broad spectrum of ceiling-approach velocities and directions. Next, we converted the sensory-motor pairs to a two-stage control policy in a continuous augmented-optical flow space. The control policy consists of a first-stage Flip-Trigger Policy, which employs a one-class support vector machine, and a second-stage Flip-Action Policy, implemented as a feed-forward neural network. To transfer the inverted-landing policy to physical systems, we utilized domain randomization and system identification techniques for a zero-shot sim-to-real transfer. As a result, we successfully achieved a range of robust inverted-landing behaviors in small quadcopters, emulating those observed in flies.

Figures (19)

• Figure 1: Bio-inspired inverted landing. An illustration comparing an example of inverted landing in a small quadcopter and a blue-bottle flyliu2019flies. Both an at-scale and a scaled-up version of the landing sequence of the blue-bottle fly are shown.
• Figure 2: Block diagram illustrating the data collection process through simulation, the training of our two-stage policy using the collected data, and the implementation of the policy in experimental tests.
• Figure 3: (a) Diagram illustrating divergent optical flow: as the quadrotor nears the ceiling, observed points radiate outward. (b) Diagram showcasing transverse optical flow, where feature points move horizontally across the field of view during the robot's translation beneath the ceiling.
• Figure 4: The figure illustrates the convergence of $\tau_{cr}$ and $a_{Rot}$ values obtained through reinforcement learning in simulation. These values are used to establish an optimal sensory-motor pair. The shaded regions indicate a distribution range of $\mu_i\pm 2\sigma_i$ for samples taken in each episode.
• Figure 5: The decision boundary formed by the One-Class SVM, implemented for the Flip Trigger Policy, is shown. This boundary distinguishes between valid and invalid triggering states for inverted landings, consistent with the policy ${\pi}_{Trg}(\textbf{s}_{Trg})$. A representative trajectory is also depicted, indicating the exact timing of the rotational maneuver trigger upon intersecting with the boundary region.