Online machine learning for calibration of autonomous earth moving vehicles
In some implementations, the EMV uses a calibration to inform autonomous control over the EMV. To calibrate an EMV, the system first selects a calibration action comprising a control signal for actuating a control surface of the EMV. Then, using a calibration model comprising a machine learning model trained based on one or more previous calibration actions taken by the EMV, the system predicts a response of the control surface to the control signal of the calibration action. After the EMV executes the control signal to perform the calibration action, the EMV system monitors the actual response of the control signal and uses that to update the calibration model based on a comparison between the predicted and monitored states of the control surface.
1. A method comprising:
training, by a training system, a machine-learned model based on a training set of data that describes movement instructions provided to one or more autonomous vehicles (AVs) and resulting AV movements, the machine-learned model configured to predict an AV movement resulting from a movement instruction and generate a control signal for actuating a control surface of the AV based on the predicted AV movement resulting from the movement instruction;
actuating a control surface of an AV to perform an actual AV movement based on a control signal;
determining, by the training system, that the actual AV movement based on the control signal is different than a predicted AV movement resulting from a movement instruction by a threshold movement amount; and
updating, by the training system, the machine-learned model such that a difference between the actual AV movement based on the control signal and the predicted AV movement meets a threshold.
2. The method of claim 1 , wherein the actual AV movement is measured based on a state of the control surface at a time when the control signal is executed.
3. The method of claim 1 , wherein the training system is configured to compare the actual AV movement to the predicted AV movement periodically.
4. The method of claim 1 , wherein the machine-learned model is further trained based on baseline training data gathered from a set of additional AVs similar to the AV.
5. The method of claim 1 , wherein the predicted AV movement comprises a predicted movement of a tool of the AV.
6. The method of claim 1 , wherein updating the machine-learned model comprises:
updating, by online learning techniques, the machine-learned model based on the actual AV movement based on the control signal.
7. The method of claim 1 , wherein the actual AV movement is compared to the predicted AV movement responsive to the AV switching between an autonomous operation mode and a manual operation mode.
8. The method of claim 1 , wherein the actual AV movement based on the control signal is determined by measuring a set of action conditions based on collected sensor data.
9. A non-transitory computer-readable storage medium storing executable instructions that, when executed by a processor, cause the processor to:
train, by a training system, a machine-learned model based on a training set of data that describes movement instructions provided to one or more autonomous vehicles (AVs) and resulting AV movements, the machine-learned model configured to predict an AV movement resulting from a movement instruction and generate a control signal for actuating a control surface of the AV based on the predicted AV movement resulting from the movement instruction;
actuate a control surface of the AV to perform an actual AV movement based on a control signal;
determine, by the training system, that the actual AV movement based on the control signal is different than a predicted AV movement resulting from a movement instruction by a threshold movement amount; and
update, by the training system, the machine-learned model such that a difference between the actual AV movement based on the control signal and the predicted AV movement meets satisfies a threshold.
10. The non-transitory computer-readable storage medium of claim 9 , wherein the actual AV movement is measured based on a state of the control surface at a time when the control signal is executed.
11. The non-transitory computer-readable storage medium of claim 9 , wherein the training system is configured to compare the actual AV movement to the predicted AV movement periodically.
12. The non-transitory computer-readable storage medium of claim 9 , wherein the machine-learned model is further trained based on baseline training data gathered from a set of additional AV similar to the AV.
13. The non-transitory computer-readable storage medium of claim 9 , wherein the predicted AV movement comprises a predicted movement of a tool of the AV.
14. The non-transitory computer-readable storage medium of claim 9 , wherein instructions for updating the machine-learned model further cause the processor to:
update, by updating, by online learning techniques, the machine-learned model based on the actual AV movement based on the control signal.
15. The non-transitory computer-readable storage medium of claim 9 , wherein the actual AV movement is compared to the predicted AV movement responsive to the AV switching between an autonomous operation mode and a manual operation mode.
16. The non-transitory computer-readable storage medium of claim 9 , wherein the actual AV movement based on the control signal is determined by measuring a set of action conditions based on collected sensor data.
17. A model training system, comprising:
a hardware processor; and
a non-transitory computer-readable storage medium storing executable instructions that, when executed by the hardware processor, cause the model training system to perform steps comprising:
training, by a training system, a machine-learned model based on a training set of data that describes movement instructions provided to one or more autonomous vehicles (AVs) and resulting AV movements, the machine-learned model configured to predict an AV movement resulting from a movement instruction and generate a control signal for actuating a control surface of the AV based on the predicted AV movement resulting from the movement instruction;
actuating a control surface of the AV to perform an actual AV movement based on a control signal;
determining, by the training system, that the actual AV movement based on the control signal for a first movement instruction is different than a predicted AV movement resulting from a movement instruction by a threshold movement amount; and
updating, by the training system, the machine-learned model such that a difference between [an] the actual AV movement based on the control signal and the predicted AV movement meets a threshold.