Learned state covariances

1. A method comprising:

determining, based at least in part on sensor data, observation data for an object detection of an object represented in the sensor data;

inputting one or more of the sensor data or the observation data associated with the object detection into a covariance machine-learning (ML) model;

receiving, as output from the covariance ML model:

an observation covariance associated with the observation data, the observation covariance representing a measure of joint variability of two or more instances of the observation data; and

model observation data;

receiving ground truth data associated with the object detection;

determining a loss based at least in part on the model observation data and the ground truth data;

altering, as a trained ML model, one or more parameters of the covariance ML model to reduce the loss; and

transmitting the trained ML model to an autonomous vehicle, the trained ML model being configured to control operation of the autonomous vehicle.

2. The method of claim 1 , wherein:

the observation data comprises one or more of:

a bounding box associated with the object;

a classification of the object;

an environment classification in which the object is located comprising one of:

a highway classification;

a residential road classification; and

a parking lot classification;

a sensor type associated with the sensor data;

distance information to the object; or

an occlusion state;

the ground truth data comprises a ground truth bounding box associated with the object; and

the model observation data comprises a model bounding box associated with the object.

3. The method of claim 1 , wherein the observation data is first observation data, the object detection is a first object detection, and the sensor data is first sensor data associated with a first sensor type, the ground truth data is first ground truth data, the model observation data is first model observation data, the observation covariance is a first observation covariance, and the first observation covariance and the model observation data are received from a first output head of the covariance ML model associated with the first sensor type, the method further comprising:

inputting second observation data associated with a second object detection of the object represented in second sensor data associated with a second sensor type into the covariance ML model;

receiving, as a second output from a second output head of the covariance ML model associated with the second sensor type:

a second observation covariance associated with the second observation data;

second model observation data;

receiving second ground truth data associated with the second object detection;

determining a second loss based at least in part on the second model observation data and the second ground truth data; and

altering, as a second trained ML model, the trained ML model to reduce the second loss.

4. The method of claim 3 , wherein:

the first output head is further associated with a first environmental classification in which the object is located; and

a third output head of the covariance ML model is associated with the first sensor type and a second environmental classification different from the first environmental classification.

5. A system comprising:

one or more processors; and

a memory storing processor-executable instructions that, when executed by the one or more processors, cause the system to perform operations comprising:

inputting, into a covariance ML model, one or more of sensor data associated with an object detection of an object represented in the sensor data or observation data for an object detection of an object represented in the sensor data;

receiving, as output from the covariance ML model, an observation covariance associated with the observation data, the observation covariance representing a measure of joint variability of two or more instances of the observation data;

determining a loss based at least in part on the output;

altering, as a trained ML model, one or more parameters of the covariance ML model to reduce the loss; and

transmitting the trained ML model to an autonomous vehicle, the trained ML model being configured to control operation of the autonomous vehicle.

6. The system of claim 5 , wherein:

the observation data comprises one or more of:

a bounding box associated with the object;

a classification of the object;

an environment classification in which the object is located comprising one of:

a highway classification;

a residential road classification; and

a parking lot classification;

a sensor type associated with the sensor data;

distance information to the object; or

an occlusion state.

7. The system of claim 5 , wherein the output from the covariance ML model further comprises model observation data comprising a model bounding box associated with the object.

8. The system of claim 7 , the operations further comprising:

receiving ground truth data associated with the object detection, wherein:

the ground truth data comprises a ground truth bounding box associated with the object.

9. The system of claim 8 , wherein:

the determining the loss is further based at least in part on the ground truth bounding box and the model bounding box; and the operations further comprising:

performing the altering of the one or more parameters by backpropagating the loss to one or more nodes of the covariance ML model, on which output the observation covariance is based at least in part.

10. The system of claim 9 , wherein the observation data is first observation data, the object detection is a first object detection, the sensor data is first sensor data associated with a first sensor type, the ground truth data is first ground truth data, the model observation data is first model observation data, the observation covariance is a first observation covariance, and the first observation covariance and the model observation data are received from a first output head of the covariance ML model associated with the first sensor type, the operations further comprising:

inputting second observation data associated with a second object detection of the object represented in second sensor data associated with a second sensor type into the covariance ML model; and

receiving, as a second output from a second output head of the covariance ML model associated with the second sensor type:

a second observation covariance associated with the second observation data; and

second model observation data.

11. The system of claim 10 , the operations further comprising:

receiving second ground truth data associated with the second object detection;

determining a second loss based at least in part on the second model observation data and the second ground truth data; and

altering, as a second trained ML model, the trained ML model to reduce the second loss.

12. The system of claim 11 , wherein:

the first output head is further associated with a first environmental classification in which the object is located; and

a third output head of the covariance ML model is associated with the first sensor type and a second environmental classification different from the first environmental classification.

13. The system of claim 11 , wherein:

the first output head is further associated with a first object classification of the object; and

a third output head of the covariance ML model is associated with the first sensor type and a second object classification different from the first object classification.

14. The system of claim 5 , wherein the covariance ML model comprises a plurality of machine learned models associated with respective sets of one or more situational factors, the one or more situational factors comprising:

an object classification of a plurality of object classifications;

an environment classification of a plurality of environment classifications in which observed objects are located comprising one of:

a highway classification;

a residential road classification; and

a parking lot classification; or

a sensor type of a plurality of sensor types.

15. One or more non-transitory computer-readable media storing processor-executable instructions that, when executed by one or more processors, cause n to perform operations comprising:

determining, based at least in part on sensor data, observation data for an object detection of an object represented in the sensor data;

inputting one or more of the sensor data or the observation data associated with the object detection into a covariance ML model;

receiving, as output from the covariance ML model an observation covariance associated with the observation data, the observation covariance representing a measure of joint variability of two or more instances of the observation data;

determining a loss based at least in part on the output;

altering, as a trained ML model, one or more parameters of the covariance ML model to reduce the loss; and

transmitting the trained ML model to an autonomous vehicle, the trained ML model being configured to control operation of the autonomous vehicle.

16. The one or more non-transitory computer-readable media of claim 15 , the operations further comprising:

receiving ground truth data associated with the object detection, wherein the ground truth data comprises a ground truth bounding box associated with the object and the determining the loss is further based at least in part on the ground truth data; and

performing the altering of the one or more parameters by backpropagating the loss.

17. The one or more non-transitory computer-readable media of claim 16 , wherein the output from the covariance ML model further comprises model observation data comprising a model bounding box associated with the object, the operations further comprising:

the determining the loss is further based at least in part on the model bounding box; and

performing the altering of the one or more parameters by backpropagating the loss to one or more nodes of the covariance ML model, on which output the observation covariance is based at least in part.

18. The one or more non-transitory computer-readable media of claim 17 , wherein the observation data is first observation data, the object detection is a first object detection, the sensor data is first sensor data associated with a first sensor type, the ground truth data is first ground truth data, the model observation data is first model observation data, the observation covariance is a first observation covariance, and the first observation covariance and the model observation data are received from a first output head of the covariance ML model associated with the first sensor type, the operations further comprising:

inputting second observation data associated with a second object detection of the object represented in second sensor data associated with a second sensor type into the covariance ML model; and

receiving, as a second output from a second output head of the covariance ML model associated with the second sensor type:

a second observation covariance associated with the second observation data; and

second model observation data.

19. The one or more non-transitory computer-readable media of claim 18 , the operations further comprising:

receiving second ground truth data associated with the second object detection;

determining a second loss based at least in part on the second model observation data and second the ground truth data; and

altering, as a second trained ML model, the trained ML model to reduce the second loss.

20. The one or more non-transitory computer-readable media of claim 15 , wherein the observation data comprises one or more of:

a bounding box for the object;

distance information to the object;

a classification of the object;

an environment classification in which the object is located comprising one of:

a highway classification;

a residential road classification; and

a parking lot classification; or

an occlusion state.