Training machine learning model for controlling autonomous vehicle
Techniques are disclosed for training one or more cost functions of an autonomous vehicle (“AV”) control system based on difference between data generated using the AV control system and manual driving data. In many implementations, manual driving data captures action(s) of a vehicle controlled by a manual driver. Additionally or alternatively, multiple AV control systems can be evaluated by comparing deviations for each AV control system, where the deviations are determined using the same set of manual driving data.
1. A method for training an autonomous vehicle control system of an autonomous vehicle, the method implemented by one or more processors and comprising:
obtaining a first instance of manual driving data corresponding to a first vehicle being driven manually by a manual driver,
wherein the first instance of manual driving data defines: a trajectory of the first vehicle controlled by a manual driver, and one or more aspects of an environment of the first vehicle while following the trajectory;
processing by the autonomous vehicle control system used to control a second vehicle, the first instance of manual driving data including the one or more aspects of the environment of the first instance of manual driving data, to generate output indicating a predicted trajectory as determined by the control system of the second vehicle;
comparing the predicted trajectory of the second vehicle with the trajectory of the first vehicle, to determine a deviation between the predicted trajectory of the second vehicle and the trajectory of the first vehicle;
identifying a statistically significant deviation between the predicted trajectory of the second vehicle with the trajectory of the first vehicle; and
determining whether the deviation satisfies one or more metrics based upon repeated similar prior instances of obtained trajectory of the first vehicle;
based on determining that the deviation satisfies the one or more metrics, generating a training instance comprising the first instance of manual driving data; and
training the autonomous vehicle control system using the generated training instance.
2. The method of claim 1 , wherein the one or more aspects of the environment of the first instance of manual driving data are processed using one or more cost functions of the autonomous vehicle control system that is used to control the second vehicle.
3. The method of claim 2 , wherein the one or more cost functions include a curb cost function that relates to controlling the autonomous vehicle with respect to a curb.
4. The method of claim 2 , wherein the one or more cost functions are represented using a machine learning model, and wherein the machine learning model processes the first instance of manual driving data as input, to output the predicted trajectory.
5. The method of claim 4 , wherein training the autonomous vehicle control system using the generated training instance comprises:
updating one or more portions of the machine learning model based on the determined deviation between the predicted trajectory of the second vehicle and the trajectory of the first vehicle.
6. The method of claim 2 , wherein the one or more cost functions include a traffic light cost function that is related to controlling the autonomous vehicle in approaching a traffic light.
7. The method of claim 2 , wherein the one or more cost functions include a cost function that is related to multiple driving conditions.
8. The method of claim 7 , wherein the cost function is related to controlling the autonomous vehicle in approaching a traffic light in rainy weather at night.
9. The method of claim 1 , wherein determining whether the deviation satisfies one or more metrics comprises using a trained machine learning deviation model.
10. The method of claim 9 , wherein the trained machine learning deviation model determines whether the deviation satisfies the one or more metrics based at least on: the deviation, the first instance of manual driving data, and the predicted trajectory.
11. The method of claim 1 wherein the predicted trajectory of the second vehicle includes at least a first measure representing jerk and a second measure representing steering angle rate.
12. The method of claim 1 , wherein the second vehicle is the autonomous vehicle or an additional autonomous vehicle.
13. The method of claim 1 , wherein the one or more aspects of the environment of the first instance of manual driving data are defined in environment data captured obtained using a sensor suite of the first vehicle.
14. The method of claim 1 , wherein a current state of the second vehicle is processed along with the one or more aspects of the environment of the first instance of manual driving data, to generate the output indicating the predicted trajectory of the second vehicle.
15. The method of claim 1 , further comprising:
deploying the trained autonomous vehicle control system in the autonomous vehicle driving on a road.
16. A method for training an autonomous vehicle control system of an autonomous vehicle, the method implemented by one or more processors and comprising:
obtaining a first instance of manual driving data corresponding to a first vehicle being driven manually by a manual driver,
wherein the first instance of manual driving data defines: a trajectory of the first vehicle controlled by a manual driver, and one or more aspects of an environment of the first vehicle while following the trajectory;
processing by the autonomous vehicle control system used to control a second autonomous vehicle, the one or more aspects of the environment of the first instance of manual driving data, to generate output indicating a predicted trajectory of the second vehicle;
comparing the predicted trajectory of the second vehicle with the trajectory of the first vehicle, to determine a deviation between the predicted trajectory of the second vehicle and the trajectory of the first vehicle;
identifying a statistically significant deviation between the predicted trajectory of the second vehicle with the trajectory of the first vehicle; and
determining whether the deviation satisfies one or more metrics based upon similar prior instances of obtained trajectory of the first vehicle;
based on determining that a deviation satisfies a predefined condition, generating a training instance comprising the first instance of manual driving data; and
training the autonomous vehicle control system using the generated training instance.
17. A method for training an autonomous vehicle control system of an autonomous vehicle, the method implemented by one or more processors and comprising:
obtaining a first instance of manual driving data in a first vehicle,
wherein the first instance of manual driving data defines: a trajectory of the first vehicle controlled by a manual driver, and one or more aspects of an environment of the first vehicle;
processing by an autonomous vehicle control system, the one or more aspects of the environment of the first instance of manual driving data, to generate a predicted trajectory of a second vehicle which is controlled by an autonomous vehicle control system;
comparing the predicted trajectory of the second vehicle with the trajectory of the first vehicle, to determine a deviation between the predicted trajectory of the second vehicle and the trajectory of the first vehicle;
identifying a statistically significant deviation between the predicted trajectory of the second vehicle with the trajectory of the first vehicle; and
determining whether the deviation satisfies one or more metrics based upon repeated similar prior instances of obtained trajectory of the first vehicle;
based on determining that the deviation is satisfies the one or more metrics, generating a training instance comprising the first instance of manual driving data; and
training the autonomous vehicle control system using the generated training instance.