Systems and methods for training predictive models for autonomous devices

1. A computer-implemented method for training a machine-learned model, the method comprising:

receiving, by a computing system comprising one or more computing devices, a rasterized image associated with a training object;

generating, by the computing system, a predicted trajectory of the training object by inputting the rasterized image into a machine-learned prediction generator model configured to output the predicted trajectory;

converting, by the computing system, the predicted trajectory into a rasterized trajectory, wherein the rasterized trajectory spatially corresponds to the rasterized image;

determining, by the computing system, a score for the predicted trajectory by inputting the rasterized trajectory and the rasterized image into a machine-learned discriminator model configured to output the score for the predicted trajectory;

determining, by the computing system, an overall loss for the machine-learned prediction generator model based, at least in part, on the score; and

training, by the computing system, the machine-learned prediction generator model by minimizing the overall loss for the machine-learned prediction generator model.

2. The computer-implemented method of claim 1 , wherein the machine-learned prediction generator model and the machine-learned discriminator model are trained end-to-end via backpropagation based on the overall loss.

3. The computer-implemented method of claim 2 , wherein the machine-learned discriminator model is trained at a discriminator learning rate and the machine-learned prediction generator model is trained at a prediction learning rate, wherein the discriminator learning rate is a higher multiple of the prediction learning rate.

4. The computer-implemented method of claim 3 , further comprising:

determining, by the computing system, the discriminator learning rate based on a comparison between a threshold network loss and the overall loss.

5. The computer-implemented method of claim 3 , wherein the discriminator learning rate is five times the prediction learning rate.

6. The computer-implemented method of claim 1 , wherein the predicted trajectory comprises a plurality of waypoints, wherein each waypoint is indicative of a future location of the training object.

7. The computer-implemented method of claim 6 , wherein converting the predicted trajectory into a rasterized trajectory comprises:

applying, by the computing system, a differentiable rasterizer to the predicted trajectory to transform the predicted trajectory into a same image space as the rasterized image.

8. The computer-implemented method of claim 7 , wherein the rasterized image comprises a plurality of pixels and each pixel of the plurality of pixels corresponds to one or more coordinates of a three-dimensional area.

9. The computer-implemented method claim 8 , wherein transforming the predicted trajectory into the same image space as the rasterized image comprises:

for each respective waypoint in the plurality of waypoints of the predicted trajectory:

determining, by the computing system, one or more coordinates of the three-dimensional area that correspond to the respective waypoint; and

assigning, by the computing system, the one or more coordinates to the respective waypoint.

10. The computer-implemented method of claim 1 , wherein the rasterized image is received from a set of training data indicative of one or more previously observed objects.

11. The computer-implemented method of claim 10 , wherein the set of training data comprises one or more ground truths indicative of a previously observed object trajectory; and

wherein the machine-learned discriminator model is previously trained to distinguish between generated trajectories and the one or more ground truths.

12. The computer-implemented method of claim 1 , wherein the rasterized image comprises data indicative of one or more features of the training object, wherein the method further comprises:

determining, by the computing system, one or more predicted control inputs for the training object by inputting the rasterized image and the predicted trajectory into a machine-learned motion model configured to output one or more predicted control inputs based, at least in part, on the data indicative of the one or more features of the training object and the predicted trajectory, wherein the score for the predicted trajectory is based, at least in part, on the one or more predicted control inputs.

13. A computing system comprising:

one or more processors; and

one or more tangible, non-transitory, computer readable media that collectively store instructions that when executed by the one or more processors cause the computing system to perform operations comprising:

obtaining a rasterized image from a set of training data;

generating a predicted trajectory for a training object by inputting the rasterized image into a machine-learned prediction generator model;

converting the predicted trajectory into a rasterized trajectory, wherein the rasterized trajectory spatially corresponds to the rasterized image;

determining a score for the predicted trajectory by inputting the rasterized trajectory and the rasterized image into a machine-learned discriminator model configured to output the score for the predicted trajectory;

determining an overall for the machine-learned prediction generator model based, at least in part, on the loss score; and

training the machine-learned prediction generator model by minimizing the overall loss for the machine-learned prediction generator model.

14. The computing system of claim 13 , wherein the rasterized image comprises data indicative of one or more features of the training object.

15. The computing system of claim 14 , further comprising:

determining one or more predicted control inputs for the training object by inputting the rasterized image and the predicted trajectory into a machine-learned motion model configured to output the one or more predicted control inputs based, at least in part, on the data indicative of the one or more features of the training object and the predicted trajectory.

16. The computing system of 15 , wherein the one or more predicted control inputs comprise at least one of a longitudinal acceleration or steering angle of the training object.

17. The computing system of claim 15 , wherein the score for the predicted trajectory is based, at least in part, on the one or more predicted control inputs.

18. The computing system of claim 12 , wherein the set of training data comprises data indicative of one or more previously observed objects.

19. A computing system, comprising:

a machine-learned prediction generator model configured to receive a rasterized image associated with a training object, and in response to receipt of the rasterized image, output a predicted trajectory of the training object;

a differential rasterizer configured to receive a predicted trajectory, and in response to receipt of the predicted trajectory, convert the predicted trajectory into a rasterized trajectory;

a machine-learned discriminator model configured to receive the rasterized trajectory and the rasterized image, and in response to receipt of the rasterized trajectory and rasterized image, output a score for the predicted trajectory;

a memory that stores a set of instructions; and

one or more processors which use the set of instructions to:

receive the rasterized image associated with the training object;

generate the predicted trajectory of the training object by inputting the rasterized image into the machine-learned prediction generator model;

convert the predicted trajectory into the rasterized trajectory by inputting the predicted trajectory into the differential rasterizer;

determine the score for the predicted trajectory by inputting the rasterized trajectory and the rasterized image into the machine-learned discriminator model;

determine an overall loss for the machine-learned prediction generator model based, at least in part, on the score; and

train the machine-learned prediction generator model by minimizing the overall loss for the machine-learned prediction generator model.

20. The computing system of claim 19 , wherein the rasterized image comprises data indicative of one or more features of the training object, wherein the computing system further comprises:

a machine-learned motion model configured to output one or more predicted control inputs based, at least in part, on the data indicative of the one or more features of the training object and the predicted trajectory; and

wherein the one or more processors are further configured to:

determine one or more predicted control inputs for the training object by inputting the rasterized image and the predicted trajectory into the machine-learned motion model, wherein the score for the predicted trajectory is based, at least in part, on the one or more predicted control inputs.