Data synthesis for autonomous control systems
An autonomous control system generates synthetic data that reflect simulated environments. Specifically, the synthetic data is a representation of sensor data of the simulated environment from the perspective of one or more sensors. The system generates synthetic data by introducing one or more simulated modifications to sensor data captured by the sensors or by simulating the sensor data for a virtual environment. The autonomous control system uses the synthetic data to train computer models for various detection and control algorithms. In general, this allows autonomous control systems to augment training data to improve performance of computer models, simulate scenarios that are not included in existing training data, and/or train computer models that remove unwanted effects or occlusions from sensor data of the environment.
1. A method of training a computer model using a set of synthesized sensor data, comprising:
obtaining a set of sensor data from one or more sensors, each instance of the sensor data including a set of sensor measurements that sense a surrounding environment of the one or more sensors;
generating, for each of a set of artifacts, the set of synthesized sensor inputs that simulate presence of the set of artifacts in the surrounding environment of the one or more sensors, the set of synthesized sensor data generated by changing the sensor measurements of at least a subset of sensor data to simulate modified responses of the one or more sensors to the presence of the set of artifacts;
training the computer model using the set of synthesized sensor data, the computer model trained to reduce differences between predictions generated based on the set of synthesized sensor data and a corresponding set of target outputs; and
applying the computer model to physical sensor data obtained from one or more physical sensors, the physical sensor data generated by sensing a surrounding environment of the one or more physical sensors having the same types of the set of artifacts that were included in the set of synthesized sensor data.
2. The method of claim 1 , wherein the set of artifacts are atmospheric conditions in the surrounding environment of the one or more sensors.
3. The method of claim 2 , wherein the computer model is configured to receive the physical sensor data and output an estimation indicating atmospheric conditions of the surrounding environment of the one or more physical sensors, and wherein the set of target outputs are labels indicating categories of the atmospheric conditions simulated in the set of synthesized sensor data.
4. The method of claim 2 , wherein the computer model is configured to receive the physical sensor data and output estimated sensor data that attenuates the atmospheric conditions sensed by the physical sensor in the physical sensor data, and wherein the set of target outputs are another subset of the sensor data that are not simulated with the atmospheric conditions.
5. The method of claim 4 , wherein the set of synthesized sensor data represent sensor data from a first sensor with a first set of sensing characteristics, and wherein the set of target outputs represent data from a second sensor with a second set of sensing characteristics different from the first set of sensing characteristics.
6. The method of claim 1 , wherein generating the set of synthesized sensor data comprises omitting at least a portion of the sensor measurements in the subset of sensor data that correspond to locations of the simulated set of artifacts in the surrounding environment of the one or more sensors.
7. The method of claim 6 , wherein the computer model is configured to receive the physical sensor data and output reconstructed sensor data that reconstructs a missing portion of the physical sensor input, and wherein the set of target outputs are sensor data that simulate sensor measurements for the omitted portion of sensor measurements that reflect sensing of the set of artifacts by the one or more sensors.
8. The method of claim 7 , wherein the computer model is further configured to receive another instance of physical sensor data that includes complementary sensor measurements for the missing portion of the physical sensor data, and wherein the reconstructed sensor data is generated based on the complementary sensor measurements for the other instance of physical sensor data.
9. The method of claim 8 , wherein the physical sensor data and the reconstructed sensor data represent sensor data from a first sensor with a first set of sensor characteristics, and the other instance of physical sensor data represent sensor data from a second sensor with a second set of sensor characteristics different from the first set of sensor characteristics.
10. The method of claim 6 , wherein the set of artifacts are at least one of a puddle or a dark-colored vehicle.
11. The method of claim 1 , wherein the computer model is applied to the physical sensor data by an autonomous control system.
12. A non-transitory computer-readable medium storing instructions for execution on a processor, the instructions when executed by the processor causing the processor to perform steps comprising:
obtaining a set of sensor data from one or more sensors, each instance of the sensor data including a set of sensor measurements that sense a surrounding environment of the one or more sensors;
generating, for each of a set of artifacts, the set of synthesized sensor inputs that simulate presence of the set of artifacts in the surrounding environment of the one or more sensors, the set of synthesized sensor data generated by changing the sensor measurements of at least a subset of sensor data to simulate modified responses of the one or more sensors to the presence of the set of artifacts;
training the computer model using the set of synthesized sensor data, the computer model trained to reduce differences between predictions generated based on the set of synthesized sensor data and a corresponding set of target outputs; and
applying the computer model to physical sensor data obtained from one or more physical sensors, the physical sensor data generated by sensing a surrounding environment of the one or more physical sensors having the same types of the set of artifacts that were included in the set of synthesized sensor data.
13. The non-transitory computer-readable medium of claim 12 , wherein the set of artifacts are atmospheric conditions in the surrounding environment of the one or more sensors.
14. The non-transitory computer-readable medium of claim 13 , wherein the computer model is configured to receive the physical sensor data and output an estimation indicating atmospheric conditions of the surrounding environment of the one or more physical sensors, and wherein the set of target outputs are labels indicating categories of the atmospheric conditions simulated in the set of synthesized sensor data.
15. The non-transitory computer-readable medium of claim 13 , wherein the computer model is configured to receive the physical sensor data and output estimated sensor data that attenuates the atmospheric conditions sensed by the physical sensor in the physical sensor data, and wherein the set of target outputs are another subset of the sensor data that are not simulated with the atmospheric conditions.
16. The non-transitory computer-readable medium of claim 15 , wherein the set of synthesized sensor data represent sensor data from a first sensor with a first set of sensing characteristics, and wherein the set of target outputs represent data from a second sensor with a second set of sensing characteristics different from the first set of sensing characteristics.
17. The non-transitory computer-readable medium of claim 12 , wherein generating the set of synthesized sensor data comprises omitting at least a portion of the sensor measurements in the subset of sensor data that correspond to locations of the simulated set of artifacts in the surrounding environment of the one or more sensors.
18. The non-transitory computer-readable medium of claim 17 , wherein the computer model is configured to receive the physical sensor data and output reconstructed sensor data that reconstructs a missing portion of the physical sensor input, and wherein the set of target outputs are sensor data that simulate sensor measurements for the omitted portion of sensor measurements that reflect sensing of the set of artifacts by the one or more sensors.
19. The non-transitory computer-readable medium of claim 18 , wherein the computer model is further configured to receive another instance of physical sensor data that includes complementary sensor measurements for the missing portion of the physical sensor data, and wherein the reconstructed sensor data is generated based on the complementary sensor measurements for the other instance of physical sensor data.
20. The non-transitory computer-readable medium of claim 19 , wherein the physical sensor data and the reconstructed sensor data represent sensor data from a first sensor with a first set of sensor characteristics, and the other instance of physical sensor data represent sensor data from a second sensor with a second set of sensor characteristics different from the first set of sensor characteristics.
21. The non-transitory computer-readable medium of claim 17 , wherein the set of artifacts are at least one of a puddle or a dark-colored vehicle.
22. The non-transitory computer-readable medium of claim 12 , wherein the computer model is applied to the physical sensor data by an autonomous control system.