Generating environmental parameters based on sensor data using machine learning
First training sensor data detected by a plurality of real-world sensors are obtained. The first training sensor data is associated with physical environment conditions. Second training sensor data detected by a plurality of virtual sensors are obtained. The second training sensor data is associated with simulated physical conditions of a virtual environment. A machine learning model is trained using both real-world and virtual training datasets including the first training sensor data, the second training sensor data, and respective sensor setting parameters of the plurality of real-world sensors and the plurality of virtual sensors. The real-world and virtual training datasets used to train the machine learning model include indications associated with the respective sensor parameter settings including one or more of the following: different scan line settings or different exposure settings. The machine learning model is provided for use in generating current parameters of an environment in which a vehicle operates.
1 . A method, comprising:
obtaining first training sensor data detected by a plurality of real-world sensors, wherein the first training sensor data is associated with physical environment conditions;
obtaining second training sensor data detected by a plurality of virtual sensors, wherein the second training sensor data is associated with simulated physical conditions of a virtual environment;
training a machine learning model using both real-world and virtual training datasets including the first training sensor data including a first sensor setting parameter of the plurality of real-world sensors, and the second training sensor data including second sensor setting parameter of the plurality of virtual sensors, wherein the first sensor setting parameter of the plurality of real-world sensors and the second sensor setting parameter of the plurality of virtual sensors are intentionally set differently to introduce controlled variations between the real-world and virtual training datasets; and
wherein the training includes: concurrently providing different data streams from the respective real-world and virtual sensor as inputs to the machine learning model; and conditioning the training of the machine learning model on the respective first and second sensor parameter settings such that the machine learning model learns to model cross-sensor dependencies by learning correlations within a same training space of the model and develops adaptability to varying sensor configurations, the machine learning model is being trained for generating a unified set of physical environment parameters in response to input of real-world sensor output data and associated real-world sensor setting parameters, and the real-world and virtual training datasets used to train the machine learning model include indications associated with the respective first and second sensor parameter settings including one or more of the following: different scan line settings, different exposure settings, or different sensor placements; and receiving map data for a geographic region, comparing mapping parameters included in the physical environment parameters generated by the machine learning model to the map data to generate error data, and applying the error data to the machine learning model; and
providing the machine learning model to a perception module of a vehicle, wherein the perception module is configured to supply generated current parameters of an environment to a planner component to generate a decision for the vehicle.
2 . The method of claim 1 , wherein training the machine learning model includes:
receiving, from a database, indications of locations of objects of a certain type in a geographic region in which one or more training vehicles operate; and
applying the indications to the machine learning model as a cost function.
3 . The method of claim 2 , wherein the objects of the certain type include traffic lights.
4 . The method of claim 1 , wherein obtaining the first training sensor data includes obtaining lidar data generated by one or more respective lidar systems.
5 . The method of claim 1 , wherein obtaining the first training sensor data includes obtaining camera data generated by one or more respective cameras operating in a visible range and/or an infrared range.
6 . The method of claim 1 , wherein obtaining the first training sensor data includes obtaining radar data generated by one or more respective radar systems.
7 . The method of claim 1 , wherein training the machine learning model includes training the machine learning model using weather data indicative of current weather conditions in the environment in which one or more training vehicles operate.
8 . The method of claim 1 , wherein training the machine learning model includes training the machine learning model for generating respective confidence scores for the physical environment parameters.
9 . The method of claim 1 , wherein training the machine learning model includes training the machine learning model for generating the physical environment parameters including an indication of a curvature of a road.
10 . The method of claim 1 , wherein training the machine learning model includes training the machine learning model to for generating the physical environment parameters including an indication of lane markings on a road.
11 . The method of claim 1 , wherein training the machine learning model includes training the machine learning model for generating the physical environment parameters including an indication of road boundaries on a road.
12 . The method of claim 1 , wherein training the machine learning model includes training the machine learning model for generating the physical environment parameters including a distance from the vehicle to an obstacle on a road.
13 . The method of claim 1 , wherein training the machine learning model incudes training at least one of a deep neural network or a convolutional neural network.
14 . A system comprising:
one or more processors configured to:
obtain first training sensor data detected by a plurality of real-world sensors, wherein the first training sensor data is associated with physical environment conditions;
obtain second training sensor data detected by a plurality of virtual sensors, wherein the second training sensor data is associated with simulated physical conditions of a virtual environment;
train a machine learning model using both real-world and virtual training datasets including the first training sensor data including a first sensor setting parameter of the plurality of real-world sensors, and the second training sensor data including second sensor setting parameter of the plurality of virtual sensors, wherein the first sensor setting parameter of the plurality of real-world sensors and the second sensor setting parameter of the plurality of virtual sensors are intentionally set differently to introduce controlled variations between the real-world and virtual training datasets; and
wherein being configured to train the machine learning model includes being configured to: concurrently provide different data streams from the respective real-world and virtual sensor as inputs to the machine learning model; and condition the training of the machine learning model on the respective first and second sensor parameter settings such that the machine learning model learns to model cross-sensor dependencies by learning correlations within a same training space of the model and develops adaptability to varying sensor configurations, the machine learning model is being trained for generating a unified set of physical environment parameters in response to input of real-world sensor output data and associated real-world sensor setting parameters, and the real-world and virtual training datasets used to train the machine learning model include indications associated with the respective first and second sensor parameter settings including one or more of the following: different scan line settings, different exposure settings, or different sensor placements; and receive map data for a geographic region, compare mapping parameters included in the physical environment parameters generated by the machine learning model to the map data to generate error data, and apply the error data to the machine learning model; and
provide the machine learning model to a perception module of a vehicle, wherein the perception module is configured to supply generated current parameters of an environment to a planner component to generate a decision for the vehicle; and
a memory coupled to at least one of the one or more processors and configured to provide the at least one of the one or more processors with instructions.
15 . The system of claim 14 , wherein being configured to train the machine learning model includes being configured to:
receive, from a database, map data for a geographic region in which one or more training vehicles operate, the map data indicating road geometry;
compare mapping parameters included in the physical environment parameters generated by the machine learning model to the map data received from the database to generate an error signal; and
apply the error signal to the machine learning model as an additional input.
16 . The system of claim 14 , wherein being configured to train the machine learning model includes being configured to:
receive, from a database, indications of locations of objects of a certain type in a geographic region in which one or more training vehicles operate; and
apply the indications to the machine learning model as a cost function.
17 . The system of claim 14 , wherein being configured to obtain the first training sensor data includes being configured to obtain lidar data generated by one or more respective lidar systems.
18 . The system of claim 14 , wherein being configured to obtain the first training sensor data includes being configured to obtain camera data generated by one or more respective cameras operating in a visible range and/or an infrared range.
19 . A computer program product, the computer program product being embodied in a non-transitory computer readable storage medium and comprising computer instructions for:
obtaining first training sensor data detected by a plurality of real-world sensors, wherein the first training sensor data is associated with physical environment conditions;
obtaining second training sensor data detected by a plurality of virtual sensors, wherein the second training sensor data is associated with simulated physical conditions of a virtual environment;
training a machine learning model using both real-world and virtual training datasets including the first training sensor data including a first sensor setting parameter of the plurality of real-world sensors, and the second training sensor data including second sensor setting parameter of the plurality of virtual sensors, wherein the first sensor setting parameter of the plurality of real-world sensors and the second sensor setting parameter of the plurality of virtual sensors are intentionally set differently to introduce controlled variations between the real-world and virtual training datasets; and
wherein the training includes: concurrently providing different data streams from the respective real-world and virtual sensor as inputs to the machine learning model; and conditioning the training of the machine learning model on the respective first and second sensor parameter settings such that the machine learning model learns to model cross-sensor dependencies by learning correlations within a same training space of the model and develops adaptability to varying sensor configurations, the machine learning model is being trained for generating physical environment parameters in response to input of real-world sensor output data and associated real-world sensor setting parameters, and the real-world and virtual training datasets used to train the machine learning model include indications associated with the respective first and second sensor parameter settings including one or more of the following: different scan line settings, different exposure settings, or different sensor placements; and receiving map data for a geographic region, comparing mapping parameters included in the physical environment parameters generated by the machine learning model to the map data to generate error data, and applying the error data to the machine learning model; and
providing the machine learning model to a perception module of a vehicle, wherein the perception module is configured to supply generated current parameters of an environment to a planner component to generate a decision for the vehicle.
20 . The computer program product of claim 19 , wherein training the machine learning model includes:
receiving, from a database, indications of locations of objects of a certain type in a geographic region in which one or more training vehicles operate; and
applying the indications to the machine learning model as a cost function.