Knowledge distillation for autonomous vehicles
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing knowledge distillation for autonomous vehicles. One of the methods includes obtaining sensor data characterizing an environment, wherein the sensor data has been captured by one or more sensors on-board a vehicle in the environment; processing, for each of one or more surrounding agents in the environment, a network input generated from the sensor data using a neural network to generate an agent discomfort prediction that characterizes a level of discomfort of the agent; combining the one or more agent discomfort predictions to generate an aggregated discomfort score; and providing the aggregated discomfort score to a path planning system of the vehicle in order to generate a future path of the vehicle.
1. A method comprising:
obtaining sensor data characterizing an environment, wherein the sensor data has been captured by one or more sensors on-board a first vehicle in the environment, wherein the sensor data comprises measures of each of one or more surrounding vehicles in the environment that are different than the first vehicle;
processing, for each of the one or more surrounding vehicles in the environment, a network input generated from the sensor data using a neural network deployed on-board the first vehicle to generate an agent discomfort prediction that characterizes a level of discomfort that the first vehicle is causing a driver or a passenger of the respective surrounding vehicle, wherein each respective surrounding vehicle is different than the first vehicle;
combining the one or more agent discomfort predictions to generate an aggregated discomfort score; and
providing the aggregated discomfort score to a path planning system of the first vehicle in order to generate a future path of the first vehicle.
2. The method of claim 1 , wherein the network input is a machine-learned network input that was learned concurrently with the training of the neural network.
3. The method of claim 1 , wherein combining the one or more agent discomfort predictions comprises one or more of:
determining a measure of central tendency of the agent discomfort predictions,
determining a minimum of the agent discomfort predictions,
determining a maximum of the agent discomfort predictions, or
processing each agent discomfort prediction using a learned function.
4. The method of claim 1 , wherein the network input for a particular surrounding agent comprises a top-down image of the environment centered on the particular surrounding agent.
5. The method of claim 1 , wherein the neural network has been trained using i) training sensor data captured by sensors on-board one or more vehicles operating in the real world and ii) user input identifying a respective comfort level of the vehicles at a plurality of time points during the operation.
6. The method of claim 1 , wherein for each surrounding agent:
the network input has been generated by processing the sensor data using a feature extractor;
the feature extractor generates the feature data using a proper subset of the sensor data; and
the proper subset of the sensor data comprises first data that characterizes one or more particular characteristics of the surrounding agent.
7. The method of claim 1 , wherein the neural network has been trained using feature distillation using a second neural network that is configured to process a second network input generated from sensor data captured by one or more sensors on-board the first vehicle and to generate a second network output characterizing a level of discomfort of the first vehicle.
8. A system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
obtaining sensor data characterizing an environment, wherein the sensor data has been captured by one or more sensors on-board a first vehicle in the environment, wherein the sensor data comprises measures of each of one or more surrounding vehicles in the environment that are different than the first vehicle;
processing, for each of one or more surrounding vehicles in the environment, a network input generated from the sensor data using a neural network deployed on-board the first vehicle to generate an agent discomfort prediction that characterizes a level of discomfort that the vehicle is causing a driver or a passenger of the respective surrounding vehicle, wherein each respective surrounding vehicle is different than the first vehicle;
combining the one or more agent discomfort predictions to generate an aggregated discomfort score; and
providing the aggregated discomfort score to a path planning system of the first vehicle in order to generate a future path of the first vehicle.
9. The system of claim 8 , wherein the network input is a machine-learned network input that was learned concurrently with the training of the neural network.
10. The system of claim 8 , wherein combining the one or more agent discomfort predictions comprises one or more of:
determining a measure of central tendency of the agent discomfort predictions,
determining a minimum of the agent discomfort predictions,
determining a maximum of the agent discomfort predictions, or
processing each agent discomfort prediction using a learned function.
11. The system of claim 8 , wherein the network input for a particular surrounding agent comprises a top-down image of the environment centered on the particular surrounding agent.
12. The system of claim 8 , wherein the neural network has been trained using i) training sensor data captured by sensors on-board one or more vehicles operating in the real world and ii) user input identifying a respective comfort level of the vehicles at a plurality of time points during the operation.
13. The system of claim 8 , wherein for each surrounding agent:
the network input has been generated by processing the sensor data using a feature extractor;
the feature extractor generates the feature data using a proper subset of the sensor data; and
the proper subset of the sensor data comprises first data that characterizes one or more particular characteristics of the surrounding agent.
14. The system of claim 8 , wherein the neural network has been trained using feature distillation using a second neural network that is configured to process a second network input generated from sensor data captured by one or more sensors on-board the first vehicle and to generate a second network output characterizing a level of discomfort of the first vehicle.
15. One or more non-transitory computer storage media encoded with computer program instructions that when executed by a plurality of computers cause the plurality of computers to perform operations comprising:
obtaining sensor data characterizing an environment, wherein the sensor data has been captured by one or more sensors on-board a first vehicle in the environment, wherein the sensor data comprises measures of each of one or more surrounding vehicles in the environment that are different than the first vehicle;
processing, for each of one or more surrounding agents in the environment, a network input generated from the sensor data using a neural network deployed on-board the first vehicle to generate an agent discomfort prediction that characterizes a level of discomfort that the first vehicle is causing a driver or a passenger of the respective surrounding vehicle, wherein each respective surrounding vehicle is different than the first vehicle;
combining the one or more agent discomfort predictions to generate an aggregated discomfort score; and
providing the aggregated discomfort score to a path planning system of the vehicle in order to generate a future path of the first vehicle.
16. The non-transitory computer storage media of claim 15 , wherein the network input is a machine-learned network input that was learned concurrently with the training of the neural network.
17. The non-transitory computer storage media of claim 15 , wherein combining the one or more agent discomfort predictions comprises one or more of:
determining a measure of central tendency of the agent discomfort predictions,
determining a minimum of the agent discomfort predictions,
determining a maximum of the agent discomfort predictions, or
processing each agent discomfort prediction using a learned function.
18. The non-transitory computer storage media of claim 15 , wherein the neural network has been trained using i) training sensor data captured by sensors on-board one or more vehicles operating in the real world and ii) user input identifying a respective comfort level of the vehicles at a plurality of time points during the operation.
19. The non-transitory computer storage media of claim 15 , wherein for each surrounding agent:
the network input has been generated by processing the sensor data using a feature extractor;
the feature extractor generates the feature data using a proper subset of the sensor data; and
the proper subset of the sensor data comprises first data that characterizes one or more particular characteristics of the surrounding agent.
20. The non-transitory computer storage media of claim 15 , wherein the neural network has been trained using feature distillation using a second neural network that is configured to process a second network input generated from sensor data captured by one or more sensors on-board the first vehicle and to generate a second network output characterizing a level of discomfort of the first vehicle.