Generating environmental parameters based on sensor data using machine learning
To generate a machine learning model for controlling autonomous vehicles, training sensor data is obtained from sensors associated with one or more vehicles, the sensor data indicative of physical conditions of an environment in which the one or more vehicles operate, and a machine learning (ML) model is trained using the training sensor data. The ML model generates parameters of the environment in response to input sensor data. A controller in an autonomous vehicle receives sensor data from one or more sensors operating in the autonomous vehicle, applies the received sensor data to the ML model to obtain parameters of an environment in which the autonomous vehicle operates, provides the generated parameters to a motion planner component to generate decisions for controlling the autonomous vehicle, and causes the autonomous vehicle to maneuver in accordance with the generated decisions.
1. A method for generating a machine learning model for controlling autonomous vehicles, the method comprising:
obtaining, by processing hardware, first training sensor data from a plurality of real-world sensors associated with one or more training vehicles, the first training sensor data being indicative of physical conditions of an environment in which the one or more training vehicles operate;
obtaining, by the processing hardware, second training sensor data from a plurality of virtual sensors, the second training sensor data being indicative of simulated physical conditions of a virtual environment; and
training, by the processing hardware, a machine learning (ML) model 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 associated with the one or more training vehicles and the plurality of virtual sensors, the ML model being trained for generating physical environment parameters in response to input of real-world sensor output data and associated real-world sensor setting parameters, wherein the real-world and virtual training datasets used to train the ML model include indications of which subsets of sensor data correspond to which of the respective sensor parameter settings including one or more of the following: different scan line distributions or different exposure settings;
wherein a controller in an autonomous vehicle (i) receives new sensor data from one or more sensors operating in the autonomous vehicle, (ii) receives respective current setting parameters from the one or more sensors of the autonomous vehicle, (iii) applies the received new sensor data and the received respective current sensor setting parameters to the ML model to generate current parameters of a current environment in which the autonomous vehicle operates, (iv) provides the generated current parameters to a motion planner component to generate decisions for controlling the autonomous vehicle, and (v) causes the autonomous vehicle to maneuver in accordance with the generated decisions.
2. The method of claim 1 , wherein training the ML model includes:
receiving, from a database, map data for a geographic region in which the one or more training vehicles operate, the map data indicating road geometry,
comparing, by the processing hardware, mapping parameters included in the physical environment parameters generated by the ML model to the map data received from the database to generate an error signal, and
applying, by the processing hardware, the error signal to the ML model as an additional input.
3. The method of claim 1 , wherein training the ML model includes:
receiving, from a database, indications of locations of objects of a certain type in a geographic region in which the one or more training vehicles operate,
applying, by the processing hardware, the indications to the ML model as a cost function.
4. The method of claim 3 , wherein the objects of the certain type include traffic lights.
5. The method of claim 1 , wherein obtaining the first training sensor data includes obtaining, from each of the one or more training vehicles, lidar data generated by one or more respective lidar systems.
6. The method of claim 1 , wherein obtaining the first training sensor data includes obtaining, from each of the one or more training vehicles, camera data generated by one or more respective cameras operating in a visible range and/or an infrared range.
7. The method of claim 1 , wherein obtaining the first training sensor data includes obtaining, from each of the one or more training vehicles, radar data generated by one or more respective radar systems.
8. The method of claim 1 , wherein training the ML model includes training the ML model using weather data indicative of current weather conditions in the environment in which the one or more training vehicles operate.
9. The method of claim 1 , wherein training the ML model includes training the ML model for generating respective confidence scores for the physical environment parameters.
10. The method of claim 1 , wherein training the ML model includes training the ML model for generating the physical environment parameters including an indication of a curvature of a road on which the autonomous vehicle travels.
11. The method of claim 1 , wherein training the ML model includes training the ML model to for generating the physical environment parameters including an indication of lane markings on a road on which the autonomous vehicle travels.
12. The method of claim 1 , wherein training the ML model includes training the ML model for generating the physical environment parameters including an indication of road boundaries on a road on which the autonomous vehicle travels.
13. The method of claim 1 , wherein training the ML model includes training the ML model for generating the physical environment parameters including a distance from the autonomous vehicle to an obstacle on a road on which the autonomous vehicle travels.
14. The method of claim 1 , wherein training the ML model incudes training at least one of a deep neural network or a convolutional neural network.
15. A computing system comprising:
processing hardware; and
a non-transitory computer-readable memory storing thereon instructions which, when executed by the processing hardware, cause the computing system to:
obtain first training sensor data from a plurality of real-world sensors associated with one or more training vehicles, the first training sensor data being indicative of physical conditions of an environment in which the one or more training vehicles operate,
obtain second training sensor data from a plurality of virtual sensors, the second training sensor data being indicative of simulated physical conditions of a virtual environment, and
train a machine learning (ML) model 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 associated with the one or more training vehicles and the plurality of virtual sensors, the ML model being trained for generating physical environment parameters in response to input of real-world sensor output data and associated real-world sensor setting parameters, wherein the real-world and virtual training datasets used to train the ML model include indications of which subsets of sensor data correspond to which of the respective sensor parameter settings including one or more of the following: different scan line distributions or different exposure settings;
wherein a controller in an autonomous vehicle (i) receives new sensor data from one or more sensors operating in the autonomous vehicle, (ii) receives respective current sensor setting parameters from the one or more sensors of the autonomous vehicle, (iii) applies the received new sensor data and the received respective current sensor setting parameters to the ML model to generate current parameters of a current environment in which the autonomous vehicle operates, (iv) provides the generated current parameters to a motion planner component to generate decisions for controlling the autonomous vehicle, and (v) causes the autonomous vehicle to maneuver in accordance with the generated decisions.
16. The computing system of claim 15 , wherein to train the ML model, the instructions cause the computing system to:
receive, from a database, map data for a geographic region in which the one or more training vehicles operate, the map data indicating road geometry,
compare mapping parameters included in the parameters generated by the ML model to the map data received from the database to generate an error signal, and
apply the error signal to the ML model as an additional input.
17. The computing system of claim 15 , wherein to train the ML model, the instructions cause the computing system to:
receive, from a database, indications of locations of objects of a certain type in a geographic region in which the one or more training vehicles operate,
apply the indications to the ML model as a cost function.
18. The computing system of claim 17 , wherein the objects of the certain type include traffic lights.
19. A controller operating in an autonomous vehicle, the controller comprising:
one or more processing units; and
a non-transitory computer-readable memory storing thereon:
a machine-learning (ML) model configured to generate physical environment parameters, in response to input of real-world sensor output data and associated real-world sensor setting parameters, wherein the ML model is trained with both real-world and virtual training datasets including first training sensor data from a plurality of real-world sensors associated with one or more training vehicles, second training sensor data from a plurality of virtual sensors, and respective sensor setting parameters of the plurality of real-world sensors and the plurality of virtual sensors, wherein the real-world and virtual training datasets used to train the ML model include indications of which subsets of sensor data correspond to which of the respective sensor parameter settings including one or more of the following: different scan line distributions or different exposure settings,
first instructions that implement a perception module configured to (i) receive new sensor data from one or more sensors operating in the autonomous vehicle, in real time, (ii) receive respective current sensor setting parameters, in real time, and (iii) apply the received new sensor data and the respective current sensor setting parameters to the ML model to generate current parameters of an environment in which the autonomous vehicle operates, and
second instructions that implement a motion planner configured to (i) receive the generated current parameters from the perception module, and (ii) generate decisions for controlling the autonomous vehicle based on the generated current parameters;
wherein the controller causes the autonomous vehicle to maneuver in accordance with the generated decisions.
20. An autonomous vehicle comprising:
vehicle maneuvering components to effectuate at least steering, acceleration, and braking of the autonomous vehicle;
one or more sensors configured to generate signals indicative of physical conditions of an environment in which the autonomous vehicle operates; and
a controller configured to (i) receive the signals from the one or more sensors, (ii) receive respective current sensor setting parameters from the one or more sensors, (iii) apply the received signals and the received respective current sensor setting parameters to a machine learning (ML) model to generate, in real time, current parameters of an environment in which the autonomous vehicle operates, (iv) apply the generated current parameters of the environment to a motion planner to generate decisions for controlling the autonomous vehicle, and (v) control the vehicle maneuvering components in accordance with the generated decisions,
wherein the ML model for generating physical environment parameters in response to input of real-world sensor output data and associated real-world sensor setting parameters is trained with both real-world and virtual training datasets including first training sensor data from a plurality of real-world sensors associated with one or more training vehicles, second training sensor data from a plurality of virtual sensors, and respective sensor setting parameters of the plurality of real-world sensors associated with the one or more training vehicles and the plurality of virtual sensors, wherein the real-world and virtual training datasets used to train the ML model include indications of which subsets of sensor data correspond to which of the respective sensor parameter settings including one or more of the following: different scan line distributions or different exposure settings.