In recent years, a variety of UAV high-altitude shows are common. In the past Chinese Valentine’s Eve night, Shanghai police also used UAVs to launch an anti fraud security propaganda.

Compared with the more mature technology of high altitude flight, how to operate UAV in a more complex and realistic low altitude environment has always been a difficulty in this field.

Recently, the joint research team of the University of technology in Lulao, Sweden and the California Institute of technology has proposed a new solution, which is expected to realize the free passage of UAVs in the real environment.

At present, this paper has been published on arXiv.

Hard core UAV can navigate and avoid obstacles

There are more and more application scenarios of UAV in real life, such as high altitude patrol, search and rescue, underground mine navigation, automatic package delivery and so on.

The rich application scenarios also put forward higher requirements for UAV automatic navigation technology. According to Bjorn Lindqvist,

“We have published several papers on UAV automatic obstacle avoidance and navigation, and in recent research, we have begun to consider how to make UAVs freely shuttle in urban environment or dynamic mobile environment without collision with people or other vehicles.”.

Bjorn Lindquist is a member of the joint research team of the University of technology in Lulao, Sweden and the California Institute of technology. Recently, the team published a new paper in IEEE robotics automation letters.

In this paper, a computing technology based on nonlinear model predictive control (NMPC) is proposed, which can provide better autonomous navigation and obstacle avoidance capability for UAVs.

More specifically, they use NMPC algorithm to predict the trajectory of obstacles in the environment around the UAV, and use classification models to distinguish different types of trajectories and predict the future position of obstacles.

Four sets of experiments were conducted to evaluate the NMPC scheme, and the results showed that the UAV model could avoid collision when surrounded by multiple moving obstacles. The four experiments were as follows:

Avoid keeping position when Pinball: the performance of UAV in keeping posture and avoiding any obstacle under different methods, in which the pinball is the obstacle.

Keeping position while avoiding pedestrians: UAV’s performance in maintaining posture and avoiding obstacles, in which pedestrians are obstacles.

Bounce condition: the obstacle provides the motion track, and detects the NMPC’s ability to predict and evade the obstacle path.

Multiple obstacles: detect the evading ability of NMPC system and test the minimum safe distance when surrounded by multiple obstacles.

“Since NMPC works by predicting and optimizing future states, this approach integrates control, local path planning and dynamic obstacle avoidance functions into a control layer, providing a fast and computationally stable solution for dynamic obstacle avoidance scenarios,” Lindqvist said. Next, we will introduce the specific experimental results.

Predictable path to avoid multiple obstacles

This paper introduces the formulation of NMPC cost function and constraints, and the method of solving dynamic obstacles. At the same time, in order to prove the effectiveness of the proposed control architecture, a variety of test experiments are carried out.

Experiment 1: keep the position when avoiding the projectile: the task of UAV is to avoid any obstacle and keep flying, and the obstacle is to launch the ball to the UAV. In this paper, NMPC constraint method is compared with other methods, such as artificial potential field, which can respond quickly and avoid obstacles in static environment.

Considering the static nature of obstacles, the space radius and obstacle radius are set to 1m respectively, which is much larger than the safety distance required to avoid static obstacles. Meanwhile, the reference of potential field controller is adjusted to be as positive as possible. The input rate constraint of NMPC is relaxed to speed up the response.

While keeping the distance from the slowly moving obstacles, the two methods can not avoid collision with the projectile obstacles, and the obstacle enters the influence area for a short time, These controllers can’t avoid the obstacles until they collide with the UAV, and they don’t know where the future position of the obstacles is. Therefore, their avoidance action may make them move along the trajectory of the obstacles.

Experiment 2: keep posture when avoiding pedestrians: in the experiment, “pedestrian” walks to UAV in the process of direct collision to test the obstacle avoidance ability and reaction speed of UAV.

The pedestrian enters the experimental barrier within 0.3 s, and the radius of the obstacle is set at 0.6 M. from the path of UAV and obstacle, it can be seen that the flight controller starts the avoidance mechanism from the time step of recognizing the trajectory.

Experiment 3: as in the first two cases, the UAV’s task is to keep the position and avoid obstacles. The radius of the obstacle is set to 0.4m, which will affect the path of UAV after the first rebound.

As shown in the figure, the time of throwing obstacles is about 0.25 s, while the response speed of the controller is 0.35 s. This shows that even the simplified trajectory model can still predict the obstacle path well enough, especially in increasing the safety radius along the prediction.

The following figure shows the predicted trajectory of the obstacle based on the initial conditions at the beginning of the avoidance operation, as well as the measurement path of the obstacle and UAV.

The UAV successfully avoided the obstacle with the minimum distance of 0.38 m, and the solver time reached the peak value of 33 Ms. Due to the unsatisfactory solver tolerance and measurement results, a small range of constraint conflicts are expected.

Experiment 4: to avoid multiple dynamic obstacles, a single UAV was set up on the collision route to avoid the UAV, and projectiles were thrown at it at the same time. The obstacle radius of both UAVs was set to 0.4. The trajectory classification and prediction scheme was applied to separate measurement of two obstacles, but in other aspects it was the same as that of a single obstacle. The trajectory of two unmanned aerial vehicles and projectiles is shown in the figure,

The minimum distance between the UAV to avoid, the nearest UAV and the obstacle is 0.45 m and 0.42 m, respectively.

It should be noted that the air avoidance UAV can keep a safe distance for a long time and avoid the incoming projectiles. In the experiment, once the obstacle UAV starts to move, the avoidance control will start immediately.

Further research direction

In general, the NMPC architecture and trajectory classification scheme proposed by the researchers have successfully provided collision free motion paths in all possible cases. The online optimization problem can be solved within the required 50 ms limit without violating established barriers or input constraints. However, this method has some limitations

The overall performance is based on the dependence on trajectory classification: even for limited trajectory studies, trajectory classification errors may occur.

Using a clear prediction of the future obstacle position: if the prediction scheme fails or the error is too large, the UAV may completely ignore the obstacles in the collision process.

It is pointed out that this work will be further optimized and expanded in the future, including more general trajectory recognition, extraction of obstacle position and velocity, optimization of trajectory classification scheme, etc. More importantly, as more obstacles expand and relate to solver time, the complexity of NMPC is analyzed to understand when it is more appropriate to solve obstacles indirectly at the control level.

Editor in charge: GT

Tagged:

Leave a Reply

Your email address will not be published.