Antoine Beyeler

Vision-based control of near-obstacle flight

This paper presents a novel control strategy, which we call optiPilot, for autonomous flight in the vicinity of obstacles. Most existing autopilots rely on a complete 6-degree-of-freedom state estimation using a GPS and an Inertial Measurement Unit (IMU) and are unable to detect and avoid obstacles. This is a limitation for missions such as surveillance and environment monitoring that may require near-obstacle flight in urban areas or mountainous environments. OptiPilot instead uses optic flow to estimate proximity of obstacles and avoid them.Our approach takes advantage of the fact that, for most platforms in translational flight (as opposed to near-hover flight), the translatory motion is essentially aligned with the aircraft main axis. This property allows us to directly interpret optic flow measurements as proximity indications. We take inspiration from neural and behavioural strategies of flying insects to propose a simple mapping of optic flow measurements into control signals that requires only a lightweight and power-efficient sensor suite and minimal processing power.In this paper, we first describe results obtained in simulation before presenting the implementation of optiPilot on a real flying platform equipped only with lightweight and inexpensive optic computer mouse sensors, MEMS rate gyroscopes and a pressure-based airspeed sensor. We show that the proposed control strategy not only allows collision-free flight in the vicinity of obstacles, but is also able to stabilise both attitude and altitude over flat terrain. These results shed new light on flight control by suggesting that the complex sensors and processing required for 6 degree-of-freedom state estimation may not be necessary for autonomous flight and pave the way toward the integration of autonomy into current and upcoming gram-scale flying platforms.

3D Vision-based Navigation for Indoor Microflyers

Fully autonomous control of ultra-light indoor airplanes has not yet been achieved because of the strong limitations on the kind of sensors that can be embedded making it difficult to obtain good estimations of altitude. We propose to revisit altitude control by considering it as an obstacle avoidance problem and introduce a novel control scheme where the ground and ceiling is avoided based on translatory optic flow, in a way similar to existing vision-based wall avoidance strategies. We show that this strategy is successful at controlling a simulated microflyer without any explicit altitude estimation and using only simple sensors and processing that have already been embedded in an existing 10-gram microflyer. This result is thus a significant step toward autonomous control of indoor flying robots.

A 10-gram vision-based flying robot

We aim at developing ultralight autonomous microflyers capable of freely flying within houses or small built environments while avoiding collisions. Our latest prototype is a fixed-wing aircraft weighing a mere 10 g, flying around 1.5 m/s, and carrying the necessary electronics for airspeed regulation and lateral collision avoidance. This microflyer is equipped with two tiny camera modules, two rate gyroscopes, an anemometer, a small microcontroller and a Bluetooth radio module. In-flight tests were carried out in a new experimentation room specifically designed for easy changing of surrounding textures.

Flying over the reality gap: From simulated to real indoor airships

Because of their ability to naturally float in the air, indoor airships (often called blimps) constitute an appealing platform for research in aerial robotics. However, when confronted to long lasting experiments such as those involving learning or evolutionary techniques, blimps present the disadvantage that they cannot be linked to external power sources and tend to have little mechanical resistance due to their low weight budget. One solution to this problem is to use a realistic flight simulator, which can also significantly reduce experimental duration by running faster than real time. This requires an efficient physical dynamic modelling and parameter identification procedure, which are complicated to develop and usually rely on costly facilities such as wind tunnels. In this paper, we present a simple and efficient physics-based dynamic modelling of indoor airships including a pragmatic methodology for parameter identification without the need for complex or costly test facilities. Our approach is tested with an existing blimp in a vision-based navigation task. Neuronal controllers are evolved in simulation to map visual input into motor commands in order to steer the flying robot forward as fast as possible while avoiding collisions. After evolution, the best individuals are successfully transferred to the physical blimp, which experimentally demonstrates the efficiency of the proposed approach.

A 10-gram Microflyer for Vision-based Indoor Navigation

We aim at developing ultralight autonomous microflyers capable of navigating within houses or small built environments. Our latest prototype is a fixed-wing aircraft weighing a mere 10 g, flying below 2 m/s and carrying the necessary electronics for airspeed regulation and obstacle avoidance. This microflyer is equipped with two tiny camera modules, two rate gyroscopes, an anemometer, a small microcontroller, and a Bluetooth radio module. In-flight tests are carried out in a new experimentation room specifically designed for easy changing of surrounding textures.

Autonomous flight at low altitude with vision-based collision avoidance and GPS-based path following

The ability to fly at low altitude while actively avoiding collisions with the terrain and other objects is a great challenge for small unmanned aircraft. This paper builds on top of a control strategy called optiPilot whereby a series of optic-flow detectors pointed at divergent viewing directions around the aircraft main axis are linearly combined into roll and pitch commands using two sets of weights. This control strategy already proved successful at controlling flight and avoiding collisions in reactive navigation experiments. This paper shows how optiPilot can be coupled with a GPS in order to provide goal-directed, nap-of-the-earth flight control in presence of static obstacles. Two fully autonomous flights of 25 minutes each are described where a 400-gram unmanned aircraft is flying at approx. 9 m above the terrain on a circular path including two copses of trees requiring efficient collision avoidance actions.

Autonomous flight at low altitude using light sensors and little computational power

The ability to fly at low altitude while actively avoiding collisions with the terrain and objects such as trees and buildings is a great challenge for small unmanned aircraft. This paper builds on top of a control strategy called optiPilot whereby a series of optic-flow detectors pointed at divergent viewing directions around the aircraft main axis are linearly combined into roll and pitch commands using two sets of weights. This control strategy already proved successful at controlling flight and avoiding collisions in reactive navigation experiments. This paper describes how optiPilot can efficiently steer a flying platform during the critical phases of hand-launched take off and landing. It then shows how optiPilot can be coupled with a GPS in order to provide goal-directed, nap-of-the-earth flight control in presence of obstacles. Two fully autonomous flights of 25 minutes each are described where a 400-gram unmanned aircraft flies at approx. 10 m above ground in a circular path including two copses of trees requiring efficient collision avoidance actions.

Optic Flow to Steer and Avoid Collisions in 3D

Optic flow is believed to be the main source of information allowing insects to control their flight. Some researchers have tried to apply this paradigm to small unmanned aerial vehicles (UAVs). So far, none of them has been able to demonstrate a fully autonomous flight of a free-flying system without relying on other cues such as GPS and/or some sort of orientation sensors (IMU, horizon detector, etc.). Getting back to the reactive approach suggested by Gibson (direct perception) and Braitenberg (direct connection from sensors to actuators), this chapter discusses how a few optic flow signals can be directly mapped into control commands for steering an aircraft in cluttered environments. The implementation of the proposed control strategy on a 10-g airplane flying autonomously in an office-sized room demonstrates how the proposed approach can result in ultra-light autopilots.

Vision-based navigation from wheels to wings

We describe an incremental approach towards the development of autonomous indoor flyers that use only vision to navigate in textured environments. In order to cope with the severe weight and energy constraints of such systems, we use spiking neural controllers that can be implemented in tiny micro-controllers and map visual information into motor commands. The network morphology is evolved by means of an evolutionary process on the physical robots. This methodology is tested in three robots of increasing complexity, from a wheeled robot to a dirigible to a winged robot. The paper describes the approach, the robots, their degrees of complexity, and summarizes results. In addition, three compatible electronic boards and a choice of vision sensors suitable for these robots are described in more details. These boards allow a comparative and gradual development of spiking neural controllers for flying robots.