Showing papers by "Xinyan Deng published in 2021"

Bio-Inspired Rapid Escape and Tight Body Flip on an At-Scale Flapping Wing Hummingbird Robot Via Reinforcement Learning

Abstract: Insects and hummingbirds are capable of acrobatic maneuvers such as rapid turns and tight 360 $^\circ$ body flips. It is challenging for bio-inspired flapping wing micro aerial vehicles to achieve animal-like performance during such maneuvers due to their limitation in mechanism sophistication and flight control. Besides being significantly underactuated compared to their natural counterparts, flight dynamics is highly nonlinear with rapidly changing, unsteady aerodynamics which remains largely unknown during aggressive maneuvers. As a result, conventional model-based control methods are inadequate to address such maneuvers effectively due to the lack of control references and aerodynamic models. In addition, during acrobatic maneuvers such as body flips, conventional control methods with underlying stabilization mechanisms would temporarily contradict the maneuvering requirements when the vehicle undergoes a full body flip including turning upside-down. In this article, reinforcement learning (RL) has been used to complement model-based control to enable animal-like maneuverability. The learned control policy serves in two different ways to either aid or even completely takeover the conventional stabilization controller in certain cases. We experimentally demonstrate animal-like maneuverability on an at-scale, dual-motor actuated flapping wing hummingbird robot. Two test cases have been performed to demonstrate the effectiveness of such integrated control methods: 1) a rapid escape maneuver recorded from hummingbirds, 2) a tight 360 $^\circ$ body flip inspired by houseflies. By leveraging RL, the hummingbird robot demonstrated a shorter completion time in escape maneuvers compared to the traditional control-based method. It also performed 360 $^\circ$ body flip successfully within only one wingspan vertical displacement.

Flying With Damaged Wings: The Effect on Flight Capacity and Bio-Inspired Coping Strategies of a Flapping Wing Robot

Abstract: Insects wings are subject to wear and tear from collisions and environmental disturbances during flight. They can tolerate both symmetrical and asymmetrical wing damages while maintaining flight capability to some extent. Drawing inspiration from nature's adaptation capabilities, we investigated the consequences of wing damage on a flapping wing micro air vehicle by quantifying the changes in wing kinematics, lift generation, control torque offset, and aerodynamic damping variations in flight tests with intact and damaged wings. For the proposed vehicle, the wing damage affected the lift generation significantly. Compared to the intact wings, the damaged ones result in increased stroke angle amplitude in order to compensate for lift loss and torque imbalance, which causes an increase in power consumption accordingly. Furthermore, asymmetric damages usually require a larger amount of additional control effort for flight stabilization compared to symmetric cases. In addition, aerodynamic damping varies as the wing areas change. All these aspects pose challenges in flight control. An adaptive controller is proposed to cope with the wing damage induced detrimental effects on flight capacity. Flight tests were conducted to validate the control performance. As a result, the robot can effectively overcome such challenges even in the case of a maximum unilateral lift loss of up to $\approx$ 22%. Such a result matches the performance of hovering hawkmoths, which can handle torque asymmetry up to 22.3 $\pm$ 7.8%. To the best of our knowledge, this is the first demonstration of FWMAVs to handle significant wing asymmetry in hover flight.

Bio-Inspired Adversarial Attack Against Deep Neural Networks

Abstract: The paper develops a new adversarial attack against deep neural networks (DNN), based on applying bio-inspired design to moving physical objects. To the best of our knowledge, this is the first work to introduce physical attacks with a moving object. Instead of following the dominating attack strategy in the existing literature, i.e., to introduce minor perturbations to a digital input or a stationary physical object, we show two new successful attack strategies in this paper. We show by superimposing several patterns onto one physical object, a DNN becomes confused and picks one of the patterns to assign a class label. Our experiment with three flapping wing robots demonstrates the possibility of developing an adversarial camouflage to cause a targeted mistake by DNN. We also show certain motion can reduce the dependency among consecutive frames in a video and make an object detector "blind", i.e., not able to detect an object exists in the video. Hence in a successful physical attack against DNN, targeted motion against the system should also be considered.

Crawl and Fly: A Bio-Inspired Robot Utilizing Unified Actuation for Hybrid Aerial-Terrestrial Locomotion

Abstract: This letter details the design and validation of a legged terrestrial locomotion system integrated onto a hummingbird-scale flapping wing robot. Two pairs of legs powered by the flight actuators allow forward, reverse, and steering motions while on the ground. Through controlled sinusoidal drive signals, this mechanically simple locomotion system allows the robot to crawl through small gaps about half the height of spaces through which it could fly thanks to the robot's transition from a vertical to horizontal orientation when crawling. On smooth hard surfaces, the robot is capable of crawling at 100 mm/s. Gaining a terrestrial mode of movement not only expands its locomotion strategies and environmental adaptability, but it also improves the endurance of the robot in particular missions due to the low power consumption. A unique design principle is that the flight components of the robot are unchanged from the original aerial robot design, and terrestrial locomotion is realized with the addition of two capstan-interlinked pairs of legs constructed from 3D printed plastic and carbon fiber rods. Based on the dual use of the flight actuators, the robot experimentally demonstrated sustained hybrid aerial-terrestrial locomotion as well as smooth crawl-to-fly transition.