Systems and methods for generating motion forecast data for a plurality of actors with respect to an autonomous vehicle
A computing system can be configured to input data that describes sensor data into an object detection model and receive, as an output of the object detection model, object detection data describing features of the plurality of the actors relative to the autonomous vehicle. The computing system can generate an input sequence that describes the object detection data. The computing system can analyze the input sequence using an interaction model to produce, as an output of the interaction model, an attention embedding with respect to the plurality of actors. The computing system can be configured to input the attention embedding into a recurrent model and determine respective trajectories for the plurality of actors based on motion forecast data received as an output of the recurrent model.
1. A computing system, comprising:
an object detection model configured to receive an input representation that describes sensor data, and in response to receipt of the input representation output object detection data describing features of a plurality of actors relative to an autonomous vehicle;
an interaction model comprising an attention model and a context aggregation model, wherein:
the attention model is configured to receive input sequence, and in response to receipt of the input sequence, generate attentional weights;
the context aggregation model is configured to receive the attentional weights and the input sequence, and in response to receipt of the attention weights and the input sequence, generate an attention embedding with respect to the plurality of actors;
a recurrent model configured to receive the attention embedding, and in response to receipt of the attention embedding, generate motion forecast data with respect to the plurality of actors, the motion forecast data describing respective trajectories for the plurality of actors;
a memory that stores a set of instructions;
one or more processors which use the set of instructions to:
input the input representation that describes the sensor data into the object detection model;
receive, as an output of the object detection model, the object detection data describing the features of the plurality of the actors relative to the autonomous vehicle;
generate an input sequence that describes the object detection data;
analyze the input sequence using an interaction model to produce, as an output of the interaction model, the attention embedding;
input the attention embedding into the recurrent model; and
determine respective trajectories for the plurality of actors based on motion forecast data received as an output of the recurrent model.
2. The computing system of claim 1 , wherein:
the interaction model is further configured to generate a relative location embedding based on the object detection data in response to receipt of the object detection data, and wherein the relative location embedding describes relative respective locations of the plurality of actors with respect to the autonomous vehicle; and
the interaction model is further configured to generate, based at least in part on the relative location embedding, the input sequence that is received by the attention model.
3. The computing system of claim 1 , wherein the context aggregation model is configured to apply the attentional weights with respect to the object detection data to generate the attention embedding.
4. The computing system of claim 1 , wherein the recurrent model comprises at least one residual block.
5. The computing system of claim 4 , wherein the recurrent model comprises at least one multi-layer perceptron.
6. The computing system of claim 5 , wherein the one or more processors further use the set of instructions to:
input the attention embedding into the at least one multi-layer perceptron of the recurrent model;
receive an output from the at least one multi-layer perceptron; and
combine the output from the least one multi-layer perceptron with the input sequence.
7. The computing system of claim 1 , wherein the one or more processors further use the set of instructions to generate the input representation that describes the sensor data by:
voxelizing the sensor data to generate voxel representations; and
augmenting the voxel representations with map data to generate the input representation.
8. The computing system of claim 1 , wherein the interaction model is configured to generate an input sequence by projecting one or more features of the object detection data to a query and a pair of key values.
9. The computing system of claim 1 , wherein the motion forecast data is indicative of respective interactions of the plurality of actors.
10. The computing system of claim 1 , wherein the respective trajectories for the plurality of actors depend on each other.
11. A computer-implement method for forecasting actor motion, the method comprising:
inputting, by a computing system comprising one or more computing devices, an input representation that describes sensor data into an object detection model that is configured to receive the input representation that describes the sensor data, and in response to receipt of the input representation, output object detection data describing features of a plurality of actors relative to an autonomous vehicle;
receiving, by the computing system and as an output of the object detection model, the object detection data describing the features of the plurality of the actors relative to the autonomous vehicle:
generating, by the computing system, an input sequence using an interaction model;
analyzing, by the computing system, the input sequence using an attention model to generate attentional weights;
receiving, by the computing system, the attentional weight and input sequence using a context aggregation model, and in response to receipt of the attentional weights and the input sequence, generate an attention embedding with respect to the plurality of actors;
inputting, by the computing system, the attention embedding into a recurrent model that is configured to receive the attention embedding, and in response to receipt of the attention embedding, generate motion forecast data with respect to the plurality of actors, the motion forecast data describing respective trajectories for the plurality of actors; and
determining, by the computing system, respective trajectories for the plurality of actors based on motion forecast data received as an output of a recurrent model, the recurrent model being configured to receive the attention embedding, and in response to receipt of the attention embedding, generate motion forecast data with respect to the plurality of actors.
12. The computer-implemented method of claim 11 , wherein:
the interaction model is further configured to generate a relative location embedding based on the object detection data in response to receipt of the object detection data, and wherein the relative location embedding describes relative respective locations of the plurality of actors with respect to the autonomous vehicle; and
the interaction model is further configured to generate, based on the relative location embedding, the input sequence that is received by the attention model.
13. The computer-implemented method of claim 11 , wherein the context aggregation model is configured to apply the attentional weights with respect to the object detection data to generate the attention embedding.
14. The computer-implemented method of claim 11 , wherein the recurrent model comprises at least one residual block.
15. The computer-implemented method of claim 14 , wherein the recurrent model comprises at least one multi-layer perceptron.
16. The computer-implemented method of claim 15 , further comprising:
inputting, by the computing system, the attention embedding into the at least one multi-layer perceptron of the recurrent model;
receiving, by the computing system, an output from the at least one multi-layer perceptron; and
combining, by the computing system, the output from the least one multi-layer perceptron with the input sequence.
17. The computer-implemented method of claim 11 , wherein the method further comprises generating, by the computing system, the input sequence.
18. The computer-implemented method of claim 11 , wherein the method further comprises:
generating, by the computing system, the input representation that describes the sensor data by voxelizing the sensor data to generate voxel representation and augmenting the voxel representations with map data to generate the input representation.
19. A computer-implement method for training one or more machine-learned systems, the method comprising:
inputting, by a computing system comprising one or more computing devices, an input representation that describes sensor data into an object detection model that is configured to receive the input representation that describes the sensor data, and in response to receipt of the input representation, output object detection data describing features of a plurality of actors relative to an autonomous vehicle;
receiving, by the computing system and as an output of the object detection model, the object detection data describing the features of the plurality of the actors relative to the autonomous vehicle;
generating, by the computing system, an input sequence using an interaction model;
analyzing, by the computing system, the input sequence using an attention model to generate attentional weights,
receiving, by the computing system, the attentional weight and input sequence using a context aggregation model, and in response to receipt of the attentional weights and the input sequence, generate an attention embedding with respect to the plurality of actors;
inputting, by the computing system, the attention embedding into a recurrent model that is configured to receive the attention embedding, and in response to receipt of the attention embedding, generate motion forecast data with respect to the plurality of actors, the motion forecast data describing respective trajectories for the plurality of actors;
determining, by the computing system, respective trajectories for the plurality of actors based on motion forecast data received as an output of a recurrent model, the recurrent model being configured to receive the attention embedding, and in response to receipt of the attention embedding, generate motion forecast data with respect to the plurality of actors; and
updating, by the computing system, one or more parameters of at least one of the object detection model, the interaction model, or the recurrent model.
20. The computer-implemented method of claim 19 , wherein updating, by the computing system, one or more parameters of the at least one of the object detection model, the interaction model, or the recurrent model comprises training, in an end-to-end configuration, at least two of the object detection model, the interaction model, or the recurrent model.