# US 2024/0174246 A1 · Learning-model predictive control with multi-step prediction for vehicle motion control

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2024/0174246 A1
- Document ID: US20240174246A1
- Status: Published application
- Kind: A1
- Application number: 18/060,023
- Filed: 2022-11-30
- Published: 2024-05-30
- Earliest claimed: 2022-11-30
- Assignees: GM Global Technology Operations LLC; University of Waterloo
- Applicants: GM Global Technology Operations LLC; University of Waterloo
- Inventors: Amir Khajepour; Chao Yu; Yubiao Zhang; Qingrong Zhao; SeyedAlireza Kasaiezadeh Mahabadi
- Web page: https://patentgrep.com/patent/US20240174246A1

## CPC

- B60W 50/10: Interpretation of driver requests or demands
- B60W 40/10: related to vehicle motion
- B60W 50/0097: {Predicting future conditions}
- B60W 60/00: Drive control systems specially adapted for autonomous road vehicles
- B60W 2050/0031: {Mathematical model of the vehicle}
- B60W 2050/0088: {Adaptive recalibration}
- B60W 2510/0604: Throttle position
- B60W 2520/28: Wheel speed
- B60W 2540/10: Accelerator pedal position
- B60W 2556/50: of positioning data, e.g. GPS [Global Positioning System] data
- B60W: CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION; CONTROL SYSTEMS SPECIALLY ADAPTED FOR HYBRID VEHICLES; ROAD VEHICLE DRIVE CONTROL SYSTEMS FOR PURPOSES NOT RELATED TO THE CONTROL OF A PARTICULAR SUB-UNIT

## Abstract

A system for learning-model predictive control (LMPC) with multi-step prediction for motion control of a vehicle includes sensors and actuators. One or more control modules each having a processor, a memory, and input/output (I/O) ports are in communication with the sensors and actuators, the processor executing program code portions stored in the memory. The program code portions cause the sensors and actuators to obtain vehicle state information, receive a driver input, and generate a desired dynamic output based on the driver input and the vehicle state information. A program code portion estimates actions of the actuators based on the vehicle state information and the driver input, and utilizes the vehicle state information, the driver input, and the estimated actions of the actuators to select one or more models of a physics-based vehicle model and a machine-learning model of the vehicle to selectively adjust commands to the actuators.

## Claims

### Claim 1

A system for learning-model predictive control (LMPC) with multi-step prediction for motion control of a vehicle, the system comprising:
one or more sensors disposed on the vehicle, the one or more of sensors measuring real-time static and dynamic data about the vehicle;
one or more actuators disposed on the vehicle, the one or more actuators altering static and dynamic characteristics of the vehicle;
one or more control modules each having a processor, a memory, and input/output (I/O) ports in communication with the one or more sensors and the one or more actuators, the processor executing program code portions stored in the memory, the program code portions comprising:
a first program code portion that causes the one or more sensors and the one or more actuators to obtain vehicle state information;
a second program code portion that receives a driver input and generates a desired dynamic output based on the driver input and the vehicle state information;
a third program code portion that estimates actions of the one or more actuators based on the vehicle state information and the driver input; and
a fourth program code portion that utilizes the vehicle state information, the driver input, and the estimated actions of the one or more actuators to select one or more models of a physics-based vehicle model and a machine-learning model of the vehicle to selectively adjust commands to the one or more actuators.

### Claim 2 (depends on 1)

The system of claim 1 wherein the one or more sensors further comprise at least one of:
inertial measurement units (IMUs) capable of measuring orientation in three dimensions, acceleration, and velocity;
Semi Active Damping Suspension (SADS) sensors capable of measuring orientation, position, velocity, acceleration in both linear and rotational aspects;
global positioning system (GPS) sensors capable of measuring a physical location of the vehicle;
wheel speed sensors;
throttle position sensors;
accelerator position sensors;
steering position sensors; and
tire pressure monitoring sensors.

### Claim 3 (depends on 1)

The system of claim 1 wherein the second program code portion receives one or more of:
a torque request from the driver; and
a steering input from the driver,
wherein based on measurements from the one or more sensors and an estimation of capacities of the one or more actuators, the torque request and the steering input from the driver are converted into a desired dynamic output that approximates vehicle dynamics that the driver input indicates.

### Claim 4 (depends on 1)

The system of claim 1 wherein the fourth program code portion further receives the vehicle state information, the driver inputs, and the estimated actions of the one or more actuators within the LMPC, wherein the LMPC includes program code for an offline training application and a real-time application.

### Claim 5 (depends on 4)

The system of claim 4 wherein the offline training application further comprises program code that, upon receiving data from the one or more sensors and from the one or more actuators:
generates a dataset;
evaluates each data point in the dataset for similarity to other data points in the dataset;
removes repeated data from the dataset; and
upon determining that a new data point in the dataset is within a predefined Euclidean distance of a previous data point in the dataset, removes the previous data point and retains the new data point in the dataset, and upon determining that a new data point in the dataset is not within a predefined Euclidean distance of a previous data point in the dataset, retains the new data point in the data set; wherein each data point in the dataset corresponds to a distinct vehicle dynamic state.

### Claim 6 (depends on 4)

The system of claim 4 wherein the real-time application further comprises:
an online machine learning process that predicts actuator outputs for current vehicle state information based on accumulated data from real-time driving, wherein predicted actuator outputs are made according to:
mean value: {tilde over (x)}k+i*|i=i, . . . , N=K*T(K+σn2I)−1; and
variance: V{tilde over (x)}k+i*|i=1, . . . , N=K**−K*T(K+σn2I)−1K* where K is a squared exponential kernel function defined as:
k
⁡
(
x
,
x
′
)
=
σ
f
2
⁢
exp
⁡
(
-
1
2
⁢
l
⁢
(
x
-
x
′
)
T
⁢
(
x
-
x
′
)
)
.

### Claim 7 (depends on 6)

The system of claim 6 wherein the real-time application further comprises program code that evaluates reliability of predicted actuator outputs based on a quantity of data points within a predefined Euclidean distance of the predicted actuator outputs and a credible interval of the predicted actuator outputs.

### Claim 8 (depends on 7)

The system of claim 7 wherein upon determining that online machine learning predicted actuator outputs are reliable, the system utilizes the machine learning model to correct predictions generated through the physics-based model of the vehicle.

### Claim 9 (depends on 8)

The system of claim 8 wherein upon determining that the online machine learning predicted actuator outputs are less reliable than actuator outputs predicted by the physics-based model, only predictions from the physics-based model are used.

### Claim 10 (depends on 6)

The system of claim 6 wherein physics-based model predictions are selectively corrected through application of a learning term Oracle:
k+i={tilde over (x)}k+i*−vk+i*|i=1, . . . N according to:
{tilde over (x)}(k+1)=Ak{tilde over (x)}(k)+Bku(k)+Dkw(k)+dk+k
{tilde over (y)}(k+1)=Ck{tilde over (x)}(k+1)
and minimizing a cost function:
J
=
∑
i
=
1
N
p
(

Y
⁡
(
k
+
i
)
-
Y
d
⁢
e
⁢
s
(
k
+
i
)

Q
2
+

u
⁡
(
k
+
i
)

R
2
+

Δ
⁢
u
⁡
(
k
+
i
)

P
2
)
where control input constraints are umin≤u≤umax and control output constraints are Ymin≤Y≤Ymax, such that when the learning term Oracle increases accuracy of the physics-based model predictions, the learning term Oracle is used to correct the physics-based model, and when the learning term Oracle does not alter or decreases accuracy of the physics-based model, only predictions of the physics-based model are used.

### Claim 11

A method for learning-model predictive control (LMPC) with multi-step prediction for motion control of a vehicle, the method comprising:
measuring, by one or more sensors disposed on the vehicle, real-time static and dynamic data about the vehicle;
altering static and dynamic characteristics of the vehicle with one or more actuators disposed on the vehicle;
utilizing one or more control modules each having a processor, a memory, and input/output (I/O) ports in communication with the one or more sensors and the one or more actuators, the processor executing program code portions stored in the memory, the program code portions comprising:
obtaining vehicle state information from the one or more sensors and the one or more actuators;
receiving a driver input and generates a desired dynamic output based on the driver input and the vehicle state information;
estimating actions of the one or more actuators based on the vehicle state information and the driver input; and
utilizing the vehicle state information, the driver input, and the estimated actions of the one or more actuators to select one or more models of a physics-based vehicle model and a machine-learning model of the vehicle to selectively adjust commands to the one or more actuators.

### Claim 12 (depends on 11)

The method of claim 11 wherein obtaining vehicle state information from the one or more sensors further comprises:
measuring orientation of the vehicle with inertial measurement units (IMUs) capable of measuring orientation in three dimensions, acceleration, and velocity;
measuring orientation, position, velocity, acceleration in both linear and rotational aspects with Semi Active Damping Suspension (SADS) sensors;
measuring a physical location of the vehicle with global positioning system (GPS) sensors;
measuring wheel speeds with wheel speed sensors;
measuring a throttle position with throttle position sensors;
measuring accelerator pedal position with accelerator position sensors;
measuring steering angle with steering position sensors; and
measuring tire pressure with tire pressure monitoring sensors.

### Claim 13 (depends on 11)

The method of claim 11 further comprising:
receiving a torque request from the driver; and
receiving a steering input from the driver,
wherein based on measurements from the one or more sensors and an estimation of capacities of the one or more actuators, the torque request and the steering input from the driver are converted into a desired dynamic output that approximates vehicle dynamics that the driver input indicates.

### Claim 14 (depends on 11)

The method of claim 11 further comprising:
receiving the vehicle state information, the driver inputs, and the estimated actions of the one or more actuators within the LMPC; and
executing program code for an offline training application and a real-time application within the LMPC.

### Claim 15 (depends on 14)

The method of claim 14 further comprising:
executing the offline training application, including:
generating a dataset;
evaluating each data point in the dataset for similarity to other data points in the dataset;
removing repeated data from the dataset; and
upon determining that a new data point in the dataset is within a predefined Euclidean distance of a previous data point in the dataset, removing the previous data point and retains the new data point in the dataset, and upon determining that a new data point in the dataset is not within a predefined Euclidean distance of a previous data point in the dataset, retaining the new data point in the data set; wherein each data point in the dataset corresponds to a distinct vehicle dynamic state.

### Claim 16 (depends on 14)

The method of claim 14 further comprising:
executing the real-time application, including:
predicting actuator outputs, with an online learning process, for current vehicle state information based on accumulated data from real-time driving, wherein predicted actuator outputs are made according to:
mean value: {tilde over (x)}k+i*|i=i, . . . , N=K*T(K+σn2I)−1; and
variance: V{tilde over (x)}k+i*|i=1, . . . , N=K**−K*T(K+σn2I)−1K*, where K is a squared exponential kernel function defined as:
k
⁡
(
x
,
x
′
)
=
σ
f
2
⁢
exp
⁡
(
-
1
2
⁢
l
⁢
(
x
-
x
′
)
T
⁢
(
x
-
x
′
)
)
.

### Claim 17 (depends on 16)

The method of claim 16 further comprising:
evaluating reliability of predicted actuator outputs based on a quantity of data points within a predefined Euclidean distance of the predicted actuator outputs and a credible interval of the predicted actuator outputs.

### Claim 18 (depends on 17)

The method of claim 17 further comprising:
utilizing the machine learning model to correct predictions generated through the physics-based model of the vehicle upon determining that the online machine learning predicted actuator outputs are reliable; and
utilizing only predictions from the physics-based model upon determining that the online machine learning predicted actuator outputs are less reliable than actuator outputs predicted by the physics-based model.

### Claim 19 (depends on 16)

The method of claim 16 further comprising:
selectively correcting the physics-based model predictions through application of a learning term Oracle:
k+i={tilde over (x)}k+i*−vk+i*|i=1, . . . N according to:
{tilde over (x)}(k+1)=Ak{tilde over (x)}(k)+Bku(k)+Dkw(k)+dk+k
{tilde over (y)}(k+1)=Ck{tilde over (x)}(k+1)
and minimizing a cost function:
J
=
∑
i
=
1
N
p
(

Y
⁡
(
k
+
i
)
-
Y
d
⁢
e
⁢
s
(
k
+
i
)

Q
2
+

u
⁡
(
k
+
i
)

R
2
+

Δ
⁢
u
⁡
(
k
+
i
)

P
2
)
where control input constraints are umin≤u≤umax and control output constraints are Ymin≤Y≤Ymax, such that when the learning term Oracle increases accuracy of the physics-based model predictions, the learning term Oracle is used to correct the physics-based model, and when the learning term Oracle does not alter or decreases accuracy of the physics-based model, only predictions of the physics-based model are used.

### Claim 20

A system for learning-model predictive control (LMPC) with multi-step prediction for motion control of a vehicle, the system comprising:
one or more sensors disposed on the vehicle, the one or more of sensors measuring real-time static and dynamic data about the vehicle;
one or more actuators disposed on the vehicle, the one or more actuators altering static and dynamic characteristics of the vehicle;
one or more control modules each having a processor, a memory, and input/output (I/O) ports in communication with the one or more sensors and the one or more actuators, the processor executing program code portions stored in the memory, the program code portions comprising:
a first program code portion that causes the one or more sensors and the one or more actuators to obtain vehicle state information;
a second program code portion that receives a driver input and generates a desired dynamic output based on the driver input and the vehicle state information, wherein the driver input includes a torque request from the driver; and
a steering input from the driver, wherein based on measurements from the one or more sensors and an estimation of capacities of the one or more actuators, the torque request and the steering input from the driver are converted into a desired dynamic output that approximates vehicle dynamics that the driver input indicates;
a third program code portion that estimates actions of the one or more actuators based on the vehicle state information and the driver input;
a fourth program code portion that utilizes the vehicle state information, the driver input, and the estimated actions of the one or more actuators to select one or more models of a physics-based vehicle model and a machine-learning model of the vehicle to selectively adjust commands to the one or more actuators, wherein the fourth program code portion receives the vehicle state information, the driver inputs, and the estimated actions of the one or more actuators within the LMPC, wherein the LMPC includes program code for an offline training application and a real-time application, wherein the offline training application further comprises program code that, upon receiving data from the one or more sensors and from the one or more actuators:
generates a dataset;
evaluates each data point in the dataset for similarity to other data points in the dataset;
removes repeated data from the dataset; and
upon determining that a new data point in the dataset is within a predefined Euclidean distance of a previous data point in the dataset, removes the previous data point and retains the new data point in the dataset, and upon determining that a new data point in the dataset is not within a predefined Euclidean distance of a previous data point in the dataset, retains the new data point in the data set; wherein each data point in the dataset corresponds to a distinct vehicle dynamic state, and
wherein the real-time application further comprises:
an online machine learning process that predicts actuator outputs for current vehicle state information based on accumulated data from real-time driving, wherein predicted actuator outputs are made according to:
mean value: {tilde over (x)}k+i*|i=i, . . . , N=K*T(K+σn2I)−1; and
variance: V{tilde over (x)}k+i*|i=1, . . . , N=K**−K*T(K+σn2I)−1K*, where K is a squared exponential kernel function defined as:
k
⁡
(
x
,
x
′
)
=
σ
f
2
⁢
exp
⁡
(
-
1
2
⁢
l
⁢
(
x
-
x
′
)
T
⁢
(
x
-
x
′
)
)
.
wherein the real-time application further comprises program code that evaluates reliability of predicted actuator outputs based on a quantity of data points within a predefined Euclidean distance of the predicted actuator outputs and a credible interval of the predicted actuator outputs, wherein upon determining that the online machine learning predicted actuator outputs are reliable, the system utilizes the machine learning model to correct predictions generated through the physics-based model of the vehicle, and upon determining that the online machine learning predicted actuator outputs are less reliable than actuator outputs predicted by the physics-based model, only predictions from the physics-based model are used, and
wherein physics-based model predictions are selectively corrected through application of a learning term Oracle:
k+i={tilde over (x)}k+i*−vk+i*|i=1, . . . N according to:
{tilde over (x)}(k+1)=Ak{tilde over (x)}(k)+Bku(k)+Dkw(k)+dk+k
{tilde over (y)}(k+1)=Ck{tilde over (x)}(k+1)
and minimizing a cost function:
J
=
∑
i
=
1
N
p
(

Y
⁡
(
k
+
i
)
-
Y
d
⁢
e
⁢
s
(
k
+
i
)

Q
2
+

u
⁡
(
k
+
i
)

R
2
+

Δ
⁢
u
⁡
(
k
+
i
)

P
2
)
where control input constraints are umin≤u≤umax and control output constraints are Ymin≤Y≤Ymax, such that when the learning term Oracle increases accuracy of the physics-based model predictions, the learning term Oracle is used to correct the physics-based model, and when the learning term Oracle does not alter or decreases accuracy of the physics-based model, only predictions of the physics-based model are used.
