Training a machine learning based model of a vehicle perception component based on sensor settings
A method for configuring a perception component of a vehicle having one or more sensors includes generating a first set of training data that includes first sensor data corresponding to a first setting of one or more sensor parameters, and an indicator of the first setting. The method also includes generating a second set of training data that includes second sensor data corresponding to a second setting of the sensor parameter(s), and an indicator of the second setting. The method further includes training the perception component, at least by training a machine learning based model using the first and second training data sets. The trained perception component is configured to generate signals descriptive of a current state of the vehicle environment by processing sensor data generated by the sensor(s), and one or more indicators indicating which setting of the sensor parameter(s) corresponds to which portions of the generated sensor data.
1. A method for configuring a perception component of a vehicle having one or more sensors configured to sense an environment through which the vehicle is moving, the method comprising:
generating, by one or more processors, a first set of training data that includes (i) first sensor data indicative of real or simulated vehicle environments, the first sensor data corresponding to a first setting of one or more sensor parameters, the first setting defining a first spatial distribution of scan lines within a point cloud, the first spatial distribution including: a uniform distribution, a sampling of a continuous mathematical distribution, or a plurality of regions each having a different uniform spatial distribution, and (ii) an indicator of the first setting;
generating, by one or more processors, a second set of training data that includes (i) second sensor data indicative of real or simulated vehicle environments, the second sensor data corresponding to a second setting of the one or more sensor parameters, the second setting defining a second spatial distribution of scan lines within a point cloud, the second spatial distribution of scan lines being different than the first spatial distribution of scan lines, and (ii) an indicator of the second setting; and
training, by one or more processors, the perception component, at least in part by training a machine learning based model of the perception component using the first and second sets of training data,
wherein the trained perception component is configured to generate signals descriptive of a current state of the environment, as the vehicle moves through the environment, by processing (i) sensor data generated by the one or more sensors, the sensor data including point clouds generated by the one or more sensors and (ii) one or more indicators indicating which setting of the one or more sensor parameters corresponds to which portions of the generated sensor data including indicating which spatial distributions correspond to which of the point clouds generated by the one or more sensors.
2. The method of claim 1 , wherein the machine learning based model consists of a single neural network.
3. The method of claim 1 , wherein the one or more sensors include one or more lidar devices.
4. The method of claim 1 , wherein the one or more sensors include one or more radar devices.
5. The method of claim 1 , wherein the sampling of the continuous mathematical distribution includes a sampling of a Gaussian distribution.
6. The method of claim 1 , further comprising utilizing, by one or more processors, the trained perception component as the vehicle moves through the environment by:
receiving first sensor data generated by a first sensor of the one or more sensors at a first time;
receiving a first indicator indicating that the received first sensor data corresponds to the first setting; and
generating, by processing the received first sensor data and the first indicator, first signals descriptive of the current state of the environment.
7. The method of claim 6 , wherein utilizing the trained perception component as the vehicle moves through the environment further includes:
receiving second sensor data generated by the first sensor, or a second sensor of the one or more sensors, at a second time;
receiving a second indicator indicating that the received second sensor data corresponds to the second setting; and
generating, by processing the received second sensor data and the second indicator, second signals descriptive of the current state of the environment.
8. The method of claim 1 , wherein the plurality of regions includes three regions comprising:
a first region covering an area of road ahead of the vehicle;
a second region covering an area that includes a horizon in front of the vehicle; and
a third region covering an area above the horizon.
9. The method of claim 1 , wherein:
the first spatial distribution is the uniform distribution of scan lines; and
the second spatial distribution includes the continuous mathematical distribution of scan lines.
10. A non-transitory computer-readable medium storing thereon instructions executable by one or more processors to implement a training procedure for training a perception component, the training procedure comprising:
generating a first set of training data that includes (i) first sensor data indicative of real or simulated vehicle environments, the first sensor data corresponding to a first setting of one or more sensor parameters, the first setting defining a first spatial distribution of scan lines within a point cloud, the first spatial distribution including: a uniform distribution, a sampling of a continuous mathematical distribution, or a plurality of regions each having a different uniform spatial distribution, and (ii) an indicator of the first setting;
generating a second set of training data that includes (i) second sensor data indicative of real or simulated vehicle environments, the second sensor data corresponding to a second setting of the one or more sensor parameters, the second setting defining a second spatial distribution of scan lines within a point cloud, the second spatial distribution of scan lines being different than the first spatial distribution of scan lines, and (ii) an indicator of the second setting; and
training the perception component, at least in part by training a machine learning based model of the perception component using the first and second sets of training data,
wherein the trained perception component is configured to generate signals descriptive of a current state of an environment, as a vehicle moves through the environment, by processing (i) sensor data generated by one or more sensors of the vehicle, the sensor data including point clouds generated by the one or more sensors and (ii) one or more indicators indicating which setting of the one or more sensor parameters corresponds to which portions of the generated sensor data including indicating which spatial distributions correspond to which of the point clouds generated by the one or more sensors.
11. The non-transitory computer-readable medium of claim 10 , wherein the machine learning based model consists of a single neural network.
12. The non-transitory computer-readable medium of claim 10 , wherein the one or more sensors include one or more lidar devices.
13. The non-transitory computer-readable medium of claim 10 , wherein the one or more sensors include one or more radar devices.
14. The non-transitory computer-readable medium of claim 10 , wherein the sampling of the continuous mathematical distribution includes a sampling of a Gaussian distribution.
15. The non-transitory computer-readable medium of claim 10 , wherein the plurality of regions includes three regions comprising:
a first region covering an area of road ahead of the vehicle;
a second region covering an area that includes a horizon in front of the vehicle; and
a third region covering an area above the horizon.
16. The non-transitory computer-readable medium of claim 10 , wherein:
the first spatial distribution is the uniform distribution of scan lines; and
the second spatial distribution includes the continuous mathematical distribution of scan lines.
17. A vehicle comprising:
one or more sensors configured to generate sensor data by sensing an environment through which the vehicle is moving, including at least a first sensor;
one or more operational subsystems; and
a computing system configured to
receive the sensor data,
generate, using a trained perception component and based on the received sensor data, signals descriptive of a current state of an environment, the perception component being trained using
a first set of training data that includes (i) first sensor data indicative of real or simulated vehicle environments, the first sensor data corresponding to a first setting of one or more sensor parameters, the first setting defining a first spatial distribution of scan lines within a point cloud, the first spatial distribution including: a uniform distribution, a sampling of a continuous mathematical distribution, or a plurality of regions each having a different uniform spatial distribution, and (ii) an indicator of the first setting, and
a second set of training data that includes (i) second sensor data indicative of real or simulated vehicle environments, the second sensor data corresponding to a second setting of the one or more sensor parameters, the second setting defining a second spatial distribution of scan lines within a point cloud, the second spatial distribution of scan lines being different than the first spatial distribution of scan lines, and (ii) an indicator of the second setting,
wherein the trained perception component is configured to generate the signals descriptive of the current state of the environment by processing (i) point clouds generated by the one or more sensors and (ii) one or more indicators indicating which spatial distributions correspond to which of the point clouds generated by the one or more sensors,
generate driving decisions based on the signals descriptive of the current state of the environment, and
cause the one or more operational subsystems to maneuver the vehicle in accordance with the generated driving decisions.
18. The vehicle of claim 17 , wherein the one or more sensors include one or more lidar devices.
19. The vehicle of claim 17 , wherein the one or more sensors include one or more radar devices.
20. The vehicle of claim 17 , wherein the sampling of the continuous mathematical distribution includes a sampling of a Gaussian distribution.
21. The vehicle of claim 17 , wherein the plurality of regions includes three regions comprising:
a first region covering an area of road ahead of the vehicle;
a second region covering an area that includes a horizon in front of the vehicle; and
a third region covering an area above the horizon.
22. The vehicle of claim 17 , wherein:
the first spatial distribution is the uniform distribution of scan lines; and
the second spatial distribution includes the continuous mathematical distribution of scan lines.