Autonomous driving system and correction learning method for autonomous driving
Provided are an autonomous driving system and a correction learning method for autonomous driving. The autonomous driving system includes a sensor configured to collect and output data required for autonomous driving, a first processor configured to output autonomous driving data on the basis of data input from the sensor, a second processor configured to output a driving data adjustment value on the basis of differences between the data input from the sensor, the autonomous driving data input from the first processor, and driving data input from driving by a human driver, and a driving part configured to perform driving on the basis of the autonomous driving data output from the first processor and the driving data adjustment value output from the second processor.
1 . An autonomous driving system comprising:
a sensor configured to collect and output data required for autonomous driving;
a first processor configured to output autonomous driving data on the basis of data input from the sensor;
a second processor configured to output a driving data adjustment value on the basis of the data input from the sensor and differences between the autonomous driving data input from the first processor and driving data input from driving by a human driver; and
a driving part configured to perform driving on the basis of the autonomous driving data output from the first processor and the driving data adjustment value output from the second processor,
wherein the second processor outputs an error spike based on a difference value between driving data input by the human driver and the autonomous driving data,
wherein the second processor comprises:
a coding part configured to express the data input from the sensor as multiple neuron spike outputs;
a state space generator configured to generate a state space on the basis of the output of the coding part;
a synapse weight processor configured to read a synapse weight from the state space generator, correct the synapse weight using an error spike based on a difference between a motion output value and a certain target value, and output the corrected synapse weight;
a motion output generator configured to generate the motion output value required for autonomous driving on the basis of the corrected synapse weight input from the synapse weight processor; and
an error generator configured to compare the motion output value output from the motion output generator with the certain target value and feed the error spike based on the difference between the motion output value and the certain target value back to the synapse weight processor.
2 . The autonomous driving system of claim 1 , wherein the difference between the autonomous driving data output from the first processor and the driving data output from the driving by the human driver is fed back to the second processor.
3 . The autonomous driving system of claim 1 , wherein the synapse weight processor reduces the synapse weight by a certain value when the error spike occurs.
4 . The autonomous driving system of claim 3 , wherein the certain value varies depending on a difference between a time at which a neuron spike occurs and a time at which the error spike occurs.
5 . The autonomous driving system of claim 1 , wherein the synapse weight processor increases the synapse weight by a certain value when the error spike does not occur.
6 . The autonomous driving system of claim 5 , wherein the certain value is a constant.
7 . The autonomous driving system of claim 1 , wherein the second processor comprises:
a suppressive output generator configured to generate a suppressive spike for suppressing a motion output on the basis of the corrected synapse weight input from the synapse weight processor;
a second synapse weight processor configured to output a corrected synapse weight using the neuron spikes input from the coding part and a suppressive spike output from the suppressive output generator,
wherein the motion output generator is configured to generate a motion output value required for autonomous driving on the basis of the corrected synapse weight input from the second synapse weight processor, and
wherein the error generator is configured to compare the motion output value output from the motion output generator with a certain target value and feed an error spike based on a difference between the motion output value and the certain target value back to the synapse weight processor.
8 . A correction learning method for autonomous driving, the correction learning method comprising:
outputting autonomous driving data;
determining whether an autonomous driving mode is active;
when the autonomous driving mode is active,
outputting an error spike based on a difference value between driving data input by a human driver and the autonomous driving data;
collecting data required for autonomous driving from a sensor;
expressing the collected data as multiple neuron spike outputs through a coding process;
generating a state space on the basis of the multiple neuron spike outputs;
reading a synapse weight from the generated state space;
correcting the synapse weight using an error spike based on a difference between a motion output value and a target value;
generating a motion output value required for autonomous driving on the basis of the corrected synapse weight;
comparing the generated motion output value with a target value and feeding the error spike based on the difference between the motion output value and the target value back to the synapse weight correction process;
calculating a driving data adjustment value on the basis of the differences between the autonomous driving data and the driving data input from driving by the human driver;
performing driving on the basis of the autonomous driving data and a driving data adjustment value; and
when the autonomous driving mode is not active, performing driving on the basis of the driving data input by the human driver.
9 . The correction learning method of claim 8 , further comprising, when the autonomous driving mode is not active, performing correction learning on the basis of the difference value between the driving data input by the human driver and the autonomous driving data.
10 . The correction learning method of claim 8 , wherein the driving data adjustment value is determined on the basis of differences between data input from a sensor, the autonomous driving data, and the driving data input from driving by the human driver.
11 . The correction learning method of claim 8 , wherein the driving data adjustment value is determined in accordance with a neuron spike for data input from a sensor and whether the error spike occurs.
12 . The correction learning method of claim 11 , wherein, when the error spike occurs, a synapse weight is reduced by a certain value.
13 . The correction learning method of claim 12 , wherein the certain value varies depending on a difference between a time at which the neuron spike occurs and a time at which the error spike occurs.
14 . The correction learning method of claim 11 , wherein, when no error spike occurs, a synapse weight is increased by a certain value.
15 . The correction learning method of claim 14 , wherein the certain value is a constant.