| CPC B60W 60/00274 (2020.02) [B60W 40/04 (2013.01); B60W 40/06 (2013.01); B60W 50/0097 (2013.01); B60W 60/0011 (2020.02); G06N 3/084 (2013.01); B60W 2554/4045 (2020.02); B60W 2554/4046 (2020.02); B60W 2554/80 (2020.02); B60W 2556/10 (2020.02); B60W 2556/35 (2020.02)] | 21 Claims |

|
1. A method performed by one or more computers, the method comprising:
obtaining scene data characterizing a scene in an environment that includes one or more agents at a current time point, the one or more agents including a target agent;
receiving, for each of a plurality of candidate future intents for the target agent after the current time point, respective intent data that characterizes the candidate future intent;
processing the scene data using a scene encoder neural network to generate an encoded representation of the scene data;
processing the respective intent data for each of the candidate future intents using a candidate encoder neural network to generate a respective encoded representation of each of the candidate future intents; and
processing the encoded representation of the scene data and the respective encoded representations of each of the candidate future intents using a predictor neural network to generate a respective trajectory prediction output that comprises a respective future trajectory prediction for each of the respective candidate future intents; and
causing an autonomous vehicle to follow a planned path using the respective prediction output.
|