Unified self-supervisory learnable vehicle motion control policy
A method includes receiving sensed vehicle-state data, actuation-command data, and surface-coefficient data from a plurality of remote vehicles, inputting the sensed vehicle-state data, the actuation-command data, and the surface-coefficient data into a self-supervised recurrent neural network (RNN) to predict vehicle states of a host vehicle in a plurality of driving scenarios, and commanding the host vehicle to move autonomously according to a trajectory determined using the vehicle states predicted using the self-supervised RNN.
1. A method for vehicle motion control, comprising:
aggregating vehicle state data from a host vehicle and a plurality of remote vehicles, wherein the vehicle state data comprises a vehicle state for each of the host vehicle and the plurality of remote vehicles, the vehicle state including velocity, position, heading, and acceleration;
normalizing the aggregated vehicle state data by chassis categories, vehicle model, and vehicle usage category for each of the host vehicle and plurality of remote vehicles;
determining a road surface condition data using sensor data from each of the host vehicle and the plurality of remote vehicles, wherein the road surface condition is a coefficient of friction between tires and a surface of a road where the host vehicle and the plurality of remote vehicles are driving;
determining an actuation command data for each of the host vehicle and the plurality of remote vehicles;
associating the vehicle state data with the road surface condition data and the actuation command data to define driving scenarios;
determining a forward dynamics for each of the host vehicle and the plurality of remote vehicles for each of the driving scenarios based on the vehicle state data, the road surface condition data, and the actuation command data, wherein the forward dynamics include the magnitude and direction of the normal forces under front tires and rear tires as well as the traction forces of the front tires and the rear tires while the host vehicle and the plurality of remote vehicles moves forward;
training a recurrent neural network using the forward dynamics for each of the host vehicle and the plurality of remote vehicles using a supervised learning method,
determining, using the recurrent neural network trained on the forward dynamics, predicted vehicle states for the host vehicle in the driving scenarios;
inputting the predicted vehicle states and the road surface condition data into the recurrent neural network to predict actuation commands for the driving scenarios; and
commanding the host vehicle to move autonomously in accordance with a trajectory that uses the actuation commands determined using the recurrent neural network.
2. The method of claim 1 , further comprising executing the recurrent neural network to predict actuation commands for the different driving scenarios, wherein executing the neural network includes determining key performance indicators using the neural network, the key performance indicators are vehicle control parameters, and determining the key performance indicators includes determining a weighted sum of an energy consumption of the host vehicle using the predicted vehicle states, a ride comfort of the host vehicle using the predicted vehicle states, and an error relative to a reference trajectory of the host vehicle using the predicted vehicle states.
3. The method of claim 2 , wherein the error relative to a reference trajectory of the host vehicle is equal to an actual yaw rate minus a commanded yaw rate.
4. The method of claim 3 , wherein the neural network includes a plurality of layers, the plurality of layers includes a batch normalization layer and a plurality of fully connected layers, and each of the plurality of fully connected layers includes a Leaky rectified linear unit (ReLU) activation function.
5. The method of claim 4 , wherein the plurality of fully connected layers includes a first fully connected layer following the batch normalization layer.
6. The method of claim 5 , wherein the plurality of layers includes a dropout layer following the first fully connected layer.
7. The method of claim 1 , wherein training the recurrent neural network includes minimizing a prediction error, wherein the prediction error is calculated using a following equation:
e
=
∑
k
=
t
-
T
+
1
t
γ
k
-
t
+
T
-
1
s
(
k
)
-
s
~
(
k
)
where:
e is the prediction error;
γ is a discounting factor, wherein 0<γ<1;
∥·∥ is Huber loss for handling outliers;
k is an index of summation;
t is an upper limit of the summation;
T is a predetermined number;
s(k) is a vehicle state of the host vehicle; and
š (k) is a predicted vehicle state of the host vehicle.
8. The method of claim 7 , further comprising training the neural network, wherein training the recurrent neural network includes fine-tuning the neural network to adapt to environmental changes, the environmental changes include tire wear changes, a presence of a replaced tire, and weight distribution changes, and training the recurrent neural network further includes determining whether the recurrent neural network has converged.
9. The method of claim 8 , further comprising:
determining whether the predicted actuation commands satisfy a plurality of safe driving envelope constraints; and
using the predicted actuation commands to control movements of the host vehicle in response to determining that the predicted actuation commands satisfy the safe driving envelope constraints, wherein the plurality of safe driving envelope constraints is defined by:
❘
"\[LeftBracketingBar]"
β
ss
❘
"\[RightBracketingBar]"
=
α
sat
+
l
r
v
x
γ
Eq
.
2
where:
α sat is a rear slip angle of the host vehicle that generates a maximum lateral force;
V x is a longitudinal axis of the host vehicle;
γ is the yaw rate of the host vehicle;
I r is the distance from a center of mass to a rear axle of the host vehicle; and
βss is a sideslip within a stable handling envelope.
10. The method of claim 8 , further comprising:
determining whether the predicted actuation commands satisfy the plurality of safe driving envelope constraints; and
refraining from using the predicted actuation commands to autonomously control movements of the host vehicle in response to determining that the predicted actuation commands do not satisfy the safe driving envelope constraints.
11. A tangible, non-transitory, machine-readable medium, comprising machine-readable instructions, that when executed by a processor, cause the processor to:
aggregate vehicle state data from a host vehicle and a plurality of remote vehicles, wherein the vehicle state data comprises a vehicle state for each of the host vehicle and the plurality of remote vehicles, the vehicle state including velocity, position, heading, and acceleration;
normalize the aggregated vehicle state data by chassis categories, vehicle model, and vehicle usage category for each of the host vehicle and plurality of remote vehicles;
determine a road surface condition data using sensor data from each of the host vehicle and the plurality of remote vehicles, wherein the road surface condition is a coefficient of friction between tires and a surface of a road where the host vehicle and the plurality of remote vehicles are driving;
determine an actuation command data for each of the host vehicle and the plurality of remote vehicles;
associate the vehicle state data with the road surface condition data and the actuation command data to define driving scenarios;
determine a forward dynamics for each of the host vehicle and the plurality of remote vehicles for each of the driving scenarios based on the vehicle state data, the road surface condition data, and the actuation command data, wherein the forward dynamics include the magnitude and direction of the normal forces under front tires and rear tires as well as the traction forces of the front tires and the rear tires while the host vehicle and the plurality of remote vehicles moves forward;
train a recurrent neural network using the forward dynamics for each of the host vehicle and the plurality of remote vehicles using a supervised learning method,
determine, using the recurrent neural network trained on the forward dynamics, predicted vehicle states for the host vehicle in the driving scenarios;
input the predicted vehicle states and the road surface condition data into the recurrent neural network to predict actuation commands for the driving scenarios; and
command the host vehicle to move autonomously in accordance with a trajectory that uses the actuation commands determined using the neural network.
12. The tangible, non-transitory, machine-readable medium of claim 11 , wherein executing the recurrent neural network includes determining key performance indicators using the neural network, the key performance indicators are vehicle control parameters, and determining the key performance indicators includes a determining a weighted sum of an energy consumption of the host vehicle using the predicted vehicle states, a ride comfort of the host vehicle using the predicted vehicle states, and an error relative to a reference trajectory of the host vehicle using the predicted vehicle states.
13. The tangible, non-transitory, machine-readable medium of claim 12 , wherein the error relative to a reference trajectory of the host vehicle is equal to an actual yaw rate minus a commanded yaw rate.
14. The tangible, non-transitory, machine-readable medium of claim 13 , wherein the neural network includes a plurality of layers, the plurality of layers includes a batch normalization layer and a plurality of fully connected layers, and each of the plurality of fully connected layers includes a Leaky rectified linear unit (ReLU) activation function.
15. The tangible, non-transitory, machine-readable medium of claim 14 , wherein the plurality of fully connected layers includes a first fully connected layer following the batch normalization layer.
16. The tangible, non-transitory, machine-readable medium of claim 15 , wherein the plurality of layers includes a dropout layer following the first fully connected layer.
17. The tangible, non-transitory, machine-readable medium of claim 16 , wherein the plurality of fully connected layers includes a second fully connected layer following the dropout layer.
18. The tangible, non-transitory, machine-readable medium of claim 17 , wherein the tangible, non-transitory, machine-readable medium, further comprising machine-readable instructions, that when executed by the processor, causes the processor to:
determine whether the predicted actuation commands satisfy a plurality of safe driving envelope constraints; and
use the predicted actuation commands to control movements of the host vehicle in response to determining that the predicted actuation commands satisfy the plurality of safe driving envelope constraints.
19. The tangible, non-transitory, machine-readable medium of claim 18 , wherein the tangible, non-transitory, machine-readable medium, further comprising machine-readable instructions, that when executed by the processor, causes the processor to:
determine whether the predicted actuation commands satisfy the plurality of safe driving envelope constraints; and
refrain from using the predicted actuation commands to control movements of the host vehicle in response to determining that the predicted actuation commands do not satisfy the plurality of safe driving envelope constraints.