Reinforcement learning-based mid-air collision avoidance
A method is provided for avoiding a conflict during a flight of an aircraft that includes a defined route of travel. The method includes receiving observations of states of the aircraft and a nearby obstacle in an environment of the aircraft as the aircraft travels the defined route. The method includes applying the states to a reinforcement learning framework to predict states of the aircraft to avoid a conflict between the aircraft and the nearby obstacle. The reinforcement learning framework determines maneuvers of the aircraft to avoid the conflict, using a policy trained using a surrogate model of the environment in which movements of the aircraft and the nearby obstacle are simulated, and determines the predicted states of the aircraft from the maneuvers. A collision avoidance trajectory is generated from the predicted states of the aircraft, and output for guidance, navigation or control of the aircraft.
1 . An apparatus for avoiding a conflict during a flight of an aircraft that includes a defined route of travel, the apparatus comprising:
a memory configured to store computer-readable program code; and
processing circuitry configured to access the memory and execute the computer-readable program code to cause the apparatus to at least:
receive observations of states of the aircraft and an obstacle in an environment of the aircraft as the aircraft travels the defined route;
apply the states to a sequential policy-optimization framework to predict and thereby determine predicted states of the aircraft to avoid a conflict between the aircraft and the obstacle, wherein the sequential policy-optimization framework is caused to:
learn a policy-optimization task by iteratively updating parameters of a trial policy using a stochastic gradient-descent optimization,
determine maneuvers of the aircraft to avoid the conflict, using a policy trained using a surrogate model of the environment in which movements of the aircraft and the obstacle are simulated, wherein initial positions of the aircraft and the obstacle are on an imaginary circle with both the aircraft and the obstacle moving toward a center of the circle, and in which the aircraft goes around the obstacle at a horizontal separation and merges back to a next safe waypoint from an original route, the surrogate model providing a sparse reward feedback for interactions including collision, returning to the original route absent collision, and failing to return to the original route absent collision, and
determine the predicted states of the aircraft from the maneuvers;
generate a collision avoidance trajectory from the predicted states of the aircraft; and
output an indication of the collision avoidance trajectory for use in at least one of guidance, navigation, or control of the aircraft.
2 . The apparatus of claim 1 , wherein the states of the aircraft and the obstacle include positions that are given as coordinates in a global geographic coordinate system, and the processing circuitry is configured to execute the computer-readable program code to cause the apparatus to further convert the coordinates to corresponding coordinates in a local coordinate system, and
wherein the states are applied to the sequential policy-optimization framework with the positions given as the corresponding coordinates in the local coordinate system.
3 . The apparatus of claim 1 , wherein the predicted states of the aircraft include positions that are given as coordinates in a local coordinate system, and the processing circuitry is configured to execute the computer-readable program code to cause the apparatus to further convert the coordinates to corresponding coordinates in a global geographic coordinate system, and
wherein the collision avoidance trajectory includes the positions given as the corresponding coordinates in the global geographic coordinate system.
4 . The apparatus of claim 1 , wherein the maneuvers that are determined include a first maneuver of the aircraft away from the defined route to avoid the conflict, and a last maneuver of the aircraft back to the defined route when the conflict has been avoided.
5 . The apparatus of claim 1 , wherein the maneuvers of the aircraft include turns described by at least an acceleration and turning rate.
6 . The apparatus of claim 1 , wherein the maneuvers of the aircraft are determined using the policy trained using the surrogate model in which the aircraft is modeled as less capable at maneuvering than the aircraft is in the environment.
7 . The apparatus of claim 1 , wherein the predicted states are determined from the maneuvers, and using the surrogate model of the environment.
8 . A method of avoiding a conflict during a flight of an aircraft that includes a defined route of travel, the method comprising:
receiving observations of states of the aircraft and an obstacle in an environment of the aircraft as the aircraft travels the defined route;
applying the states to a sequential policy-optimization framework to predict and thereby determine predicted states of the aircraft to avoid a conflict between the aircraft and the obstacle, wherein the sequential policy-optimization framework:
learns a policy-optimization task by iteratively updating parameters of a trial policy using a stochastic gradient-descent optimization,
determines maneuvers of the aircraft to avoid the conflict, using a policy trained using a surrogate model of the environment in which movements of the aircraft and the obstacle are simulated, wherein initial positions of the aircraft and the obstacle are on an imaginary circle with both the aircraft and the obstacle moving toward a center of the circle, and in which the aircraft goes around the obstacle at a horizontal separation and merges back to a next safe waypoint from an original route, the surrogate model providing a sparse reward feedback for interactions including collision, returning to the original route absent collision, and failing to return to the original route absent collision, and
determines the predicted states of the aircraft from the maneuvers;
generating a collision avoidance trajectory from the predicted states of the aircraft; and
outputting an indication of the collision avoidance trajectory for use in at least one of guidance, navigation or control of the aircraft.
9 . The method of claim 8 , wherein the states of the aircraft and the obstacle include positions that are given as coordinates in a global geographic coordinate system, and the method further comprises converting the coordinates to corresponding coordinates in a local coordinate system, and
wherein the states are applied to the sequential policy-optimization framework with the positions given as the corresponding coordinates in the local coordinate system.
10 . The method of claim 8 , wherein the predicted states of the aircraft include positions that are given as coordinates in a local coordinate system, and the method further comprises converting the coordinates to corresponding coordinates in a global geographic coordinate system, and
wherein the collision avoidance trajectory includes the positions given as the corresponding coordinates in the global geographic coordinate system.
11 . The method of claim 8 , wherein the maneuvers that are determined include a first maneuver of the aircraft away from the defined route to avoid the conflict, and a last maneuver of the aircraft back to the defined route when the conflict has been avoided.
12 . The method of claim 8 , wherein the maneuvers of the aircraft include turns described by at least an acceleration and turning rate.
13 . The method of claim 8 , wherein the maneuvers of the aircraft are determined using the policy trained using the surrogate model in which the aircraft is modeled as less capable at maneuvering than the aircraft is in the environment.
14 . The method of claim 13 , wherein the aircraft is modeled in the surrogate model as a dynamic model with parameters that describe operational capabilities including maneuverability of the aircraft in the surrogate model, and at least one of the parameters is limited to a subset of a range that describes maneuverability of the aircraft in the environment.
15 . The method of claim 8 , wherein the maneuvers of the aircraft are determined using the policy that is implemented as a fully connected deep neural network.
16 . The method of claim 8 , wherein the predicted states are determined from the maneuvers using a transition model that is separate and independent from the surrogate model.
17 . The method of claim 8 , wherein learning the policy-optimization task comprises learning as a Markov Decision problem comprising Markovian states, transitions, and transition rewards.
18 . The method of claim 8 , wherein learning the policy-optimization task comprises learning according to an actor-critic model.
19 . A computer-readable storage medium for avoiding a conflict during a flight of an aircraft that includes a defined route of travel, the computer-readable storage medium being non-transitory and having computer-readable program code stored therein that, in response to execution by processing circuitry, causes an apparatus to at least:
receive observations of states of the aircraft and an obstacle in an environment of the aircraft as the aircraft travels the defined route;
apply the states to a sequential policy-optimization framework to predict and thereby determine predicted states of the aircraft to avoid a conflict between the aircraft and the obstacle, wherein the sequential policy-optimization framework caused to:
learn a policy-optimization task by iteratively updating parameters of a trial policy using a stochastic gradient-descent optimization,
determine maneuvers of the aircraft to avoid the conflict, using a policy trained using a surrogate model of the environment in which movements of the aircraft and the obstacle are simulated, wherein initial positions of the aircraft and the obstacle are on an imaginary circle with both the aircraft and the obstacle moving toward a center of the circle, and in which the aircraft goes around the obstacle at a horizontal separation and merges back to a next safe waypoint from an original route, and
determine the predicted states of the aircraft from the maneuvers;
generate a collision avoidance trajectory from the predicted states of the aircraft; and
output an indication of the collision avoidance trajectory for use in at least one of guidance, navigation or control of the aircraft.
20 . The computer-readable storage medium of claim 19 , wherein the surrogate model provides sparse reward feedback with different rewards for interactions, including:
(a) collision;
(b) returning to the original route absent collision; and
(c) failing to return to the original route absent collision.