engineers used deep learning methods to teach the drones to force a path through space with
obstacles, avoiding collisions between drones, while taking into account aerodynamic
disturbances generated by drones in the swarm. In comparison with the traditional algorithm for optimal
mutual avoidance of collisions, a new algorithm of path planning was on
20 percent more efficient in most experimental situations, without requiring large computational resources and can cope even with such complex situations as the swapping of two drones in a narrow corridor. A description of the algorithms published in the article in the journal IEEE Robotics and Automation Letters, Preprint available at arXiv.org.
In the modern world swarms of drones are widely used only in the field
entertainment, however, there are more tasks
for which you can use a group of flying robots. For example, for the problem
search and rescue of victims in emergency situations instead of a single
robots are much better suited swarm of small drones, which can quickly
to survey a large area of damaged buildings, debris or other area.
For the effective control of the swarm and control the interaction between the individual
drones need to solve two key problems. First, the problem
navigation and path planning for each robot with a lack of information about the
space: in a constantly changing environment in addition to stationary obstacles
there are other drones from the swarm, which clashes with the need
to avoid. Secondly, the multicopter create air flow that you want
to consider when close traffic drones, so they don’t blow each other
Engineers from the California Institute of technology
under the leadership of soon-Jo Chung (Soon-Jo Chung) from laboratory of jet movement NASA have used deep learning algorithms to solve these tasks. The first algorithm is called GLAS (Global-to-Local Autonomy Synthesis) and is responsible for navigation
and allows you to lead the way in a complex, dynamic environment. Second, Neural-Swarm, is responsible for the adjustment of the trajectory taking into account aerodynamic disturbances generated by nearby drones.
As a platform for experiments, the engineers used lightweight (34 grams) quadcopter Crazyflie
2.0marked with markers for tracking. To calculate its trajectory of each drone uses
only information on the situation in its local environment fixed
obstacles and other drones, without having ideas about the global picture in
General. However, in order to train a neural network which solves the problem of searching
way engineers first used global planner trajectories.
They created a simulation model that represents the virtual space of 64 square meters, which with the help of the global scheduler are generated in increments of 0.5 seconds of the trajectory of simultaneous movement of different numbers of drones (4, 8 or 16), it also varies the number and location of obstacles. Thereafter the obtained data array
retrieves information about the relative position of obstacles only in the local environment of a given radius around each drone and the trajectory of their movement in each moment of time, which are generated by the global planner trajectories.
The total data volume for all options environment approximately 40 million points. This data set observation-action is then used for training a neural network with ReLU activation function and architecture, based on the approach of the Deep Sets, which is of vectors of fixed dimension is a functional defined on sets and invariant to permutations. The choice of this approach is associated with considerable fluctuations of the dimension of the vector of observations of each drone, because at any time the number of obstacles and other drones in scope can change dramatically.
it turned out, this approach of training local scheduler on data generated in simulations using the global scheduler, allows to reduce the difficulty of the drones close to obstacles due to problems of local minima inherent in decentralized algorithms build the path as a result of incomplete information
about the environment.
The experiments showed that in most scenarios the new algorithm GLAS 20 percent more efficient than traditionally used in such tasks, the algorithm of optimal reciprocal collision avoidance ORCA (Optimal Reciprocal Collision Avoidance). As the index of effectiveness is estimated the number of drones that successfully reached the destination point without colliding with each other and not falling into the trap of local minimum.
In addition, GLAS, being a decentralized algorithm that allows to scale the number of drones in the swarm, and low requirements to computing resources provide an opportunity to use a cheap microcontroller on-Board the drone. Calculations in the case of a single drone-neighbor scope is about 3.4 milliseconds. When you increase the number of drones up to three this time increases to five milliseconds, which allows the use of available on-Board computational resources for calculations in real time with a frequency of 40 Hz.
To solve the problem of aerodynamic interaction,
engineers used data on a random close overflights by drones when they have an impact on the trajectory of each other
air flow from the rotors. For this, drones were moved within a defined space according to the random routes — the computer randomly choose
the target points for each drone fixed
frequency. The number of drones in the experiments ranged from two to four. As
algorithm for avoiding collisions in this case, for simplicity, we used the method of artificial potentials: target points had
attractive force, whereas the neighboring drones pushed each other.
The collected data about the state of drones in each moment of the experiment (their
the position in space, velocity, acceleration, thrust) and calculated their
using the z-component
disturbing forces are then used for training a neural network constructed
based on the approach of Deep Sets with the ReLU activation function. The final algorithm with the trained parameters
allowed on the fly to change the rotors to compensate for the perturbation
the trajectory flown from nearby drones. According to the authors a new approach
reduced error motion in the vertical plane due to the aerodynamic interaction by almost four times.
Task management and
interaction is important not only for flying but for ground teams of robots. So, the University
Lincoln proposed system with open source code to control a swarm
robots with the “light of the pheromones”.