Systems and methods for training predictive models for autonomous devices
Systems and methods for training machine-learned models are provided. A method can include receiving a rasterized image associated with a training object and generating a predicted trajectory of the training object by inputting the rasterized image into a first machine-learned model. The method can include converting the predicted trajectory into a rasterized trajectory that spatially corresponds to the rasterized image. The method can include utilizing a second machine-learned model to determine an accuracy of the predicted trajectory based on the rasterized trajectory. The method can include determining an overall loss for the first machine-learned model based on the accuracy of the predictive trajectory as determined by the second machine-learned model. The method can include training the first machine-learned model by minimizing the overall loss for the first machine-learned model.
1. An autonomous vehicle control system for controlling an autonomous vehicle, the autonomous vehicle control system comprising:
one or more processors; and
one or more tangible, non-transitory, computer-readable media that store instructions that are executable to cause the autonomous vehicle control system to perform operations, the operations comprising:
generating, based at least in part on sensor data descriptive of an object in an environment of the autonomous vehicle, and using a machine-learned prediction generator, a predicted trajectory of the object,
wherein the machine-learned prediction generator was trained based at least in part on a score, the score generated based at least in part on a prediction by a machine-learned discriminator, the prediction indicating whether a training trajectory processed by the machine-learned discriminator was generated by the machine-learned prediction generator; and
initiating control of the autonomous vehicle based at least in part on the predicted trajectory of the object.
2. The autonomous vehicle control system of claim 1 , wherein the training trajectory was rasterized for processing by the machine-learned discriminator.
3. The autonomous vehicle control system of claim 2 , wherein the training trajectory was rasterized to spatially correspond to a coordinate space of the sensor data.
4. The autonomous vehicle control system of claim 3 , wherein the training trajectory was generated by a differentiable rasterizer, and wherein the machine-learned prediction generator was trained based at least in part on a loss backpropagated through the differentiable rasterizer, the loss determined based at least in part on the score.
5. The autonomous vehicle control system of claim 1 , wherein the machine-learned discriminator was trained to discriminate between:
trajectories generated by the machine-learned prediction generator, and
ground truth trajectories.
6. The autonomous vehicle control system of claim 1 , wherein the machine-learned discriminator was pre-trained before training of the machine-learned prediction generator.
7. The autonomous vehicle control system of claim 1 , wherein the machine-learned prediction generator and the machine-learned discriminator were iteratively trained based on an overall loss, wherein:
in a first cycle, one or more parameters of the machine-learned prediction generator were frozen while training the machine-learned discriminator to increase a loss, and
in a second cycle, one or more parameters of the machine-learned discriminator were frozen while training the machine-learned prediction generator to decrease the loss.
8. The autonomous vehicle control system of claim 7 , wherein the machine-learned prediction generator was trained with a first learning rate, and wherein the machine-learned discriminator was trained with a second, faster learning rate.
9. The autonomous vehicle control system of claim 8 , wherein the second cycle was performed a greater number of times than the first cycle.
10. One or more tangible, non-transitory, computer-readable media that store instructions that are executable to cause one or more processors to perform operations, the operations comprising:
generating, based at least in part on sensor data descriptive of an object in an environment of an autonomous vehicle, and using a machine-learned prediction generator, a predicted trajectory of the object,
wherein the machine-learned prediction generator was trained based at least in part on a score, the score generated based at least in part on a prediction by a machine-learned discriminator, the prediction indicating whether a training trajectory processed by the machine-learned discriminator was generated by the machine-learned prediction generator; and
initiating control of the autonomous vehicle based at least in part on the predicted trajectory of the object.
11. The one or more tangible, non-transitory, computer-readable media of claim 10 , wherein the training trajectory was rasterized for input to the machine-learned discriminator.
12. The one or more tangible, non-transitory, computer-readable media of claim 11 , wherein the training trajectory was rasterized to spatially correspond to a coordinate space of the sensor data.
13. The one or more tangible, non-transitory, computer-readable media of claim 12 , wherein the training trajectory was generated by a differentiable rasterizer, and wherein the machine-learned prediction generator was trained based at least in part on a loss backpropagated through the differentiable rasterizer, the loss determined based at least in part on the score.
14. The one or more tangible, non-transitory, computer-readable media of claim 10 , wherein the machine-learned discriminator was trained to discriminate between:
trajectories generated by the machine-learned prediction generator, and
ground truth trajectories.
15. The one or more tangible, non-transitory, computer-readable media of claim 10 , wherein the machine-learned discriminator was pre-trained before training of the machine-learned prediction generator.
16. The one or more tangible, non-transitory, computer-readable media of claim 10 , wherein the machine-learned discriminator was pre-trained before training of the machine-learned prediction generator.
17. The one or more tangible, non-transitory, computer-readable media of claim 10 , wherein the machine-learned prediction generator and the machine-learned discriminator were iteratively trained based on an overall loss, wherein:
in a first cycle, one or more parameters of the machine-learned prediction generator were frozen while training the machine-learned discriminator to increase a loss, and
in a second cycle, one or more parameters of the machine-learned discriminator were frozen while training the machine-learned prediction generator to decrease the loss.
18. The one or more tangible, non-transitory, computer-readable media of claim 17 , wherein the machine-learned prediction generator was trained with a first learning rate, and wherein the machine-learned discriminator was trained with a second, faster learning rate.
19. The one or more tangible, non-transitory, computer-readable media of claim 18 , wherein the second cycle was performed a greater number of times than the first cycle.
20. A computer-implemented method for end-to-end training a machine-learned prediction generator for an autonomous vehicle control system, the method comprising:
generating, based at least in part on data descriptive of an object in an environment, and using the machine-learned prediction generator, a predicted trajectory of the object;
determining, using a machine-learned discriminator, a score for the predicted trajectory, the score generated based at least in part on a prediction by a machine-learned discriminator, the prediction indicating whether the predicted trajectory processed by the machine-learned discriminator was generated by the machine-learned prediction generator;
backpropagating a loss through the machine-learned discriminator and the machine-learned prediction generator, the loss based at least in part on the score; and
updating one or more parameters of the machine-learned prediction generator based at least in part on the loss.