Autonomous vehicle planning and prediction

1 . A computer-implemented method of probabilistically predicting a target location of an external actor, the method comprising:

receiving, at a computer, sensor inputs for detecting and tracking the external actor;

applying object tracking to the sensor inputs, in order to track the external actor, and thereby determine an observed trace of the external actor over a time interval from a first time to a third time;

determining based on a map a set of available target locations for the external actor;

for each target location of the set of available target locations, generating a respective first trajectory model over said time interval from the first time to the third time, based on a location of the external actor at the first time, and the respective target location on the map, wherein each first trajectory model includes one or more first trajectories from the location of the external actor at the first time to the respective target location;

for each target location, determining, a likelihood that the external actor is attempting to reach the respective target location based on (1) the first trajectory model for the respective target location and (2) an observed trace of the external actor between the first time and a second time, wherein the second time is between the first time and the third time; and

generating one or more control signals based on the determined likelihood, and providing the one or more control signals to a drive mechanism of an autonomous vehicle.

2 . The method of claim 1 , implemented in the autonomous vehicle, wherein a planner of the autonomous vehicle makes an autonomous driving decision in dependence on the likelihood that the external actor is attempting to reach at least one target location of the set of available target locations and on the sensor inputs obtained using a sensor system of the autonomous vehicle, and wherein the generating of the one or more control signals is further based on the autonomous driving decision.

3 . The method of claim 1 , wherein the one or more first trajectories of the first trajectory model for each target location include:

a single predicted trajectory associated with the respective target location, or

a distribution of predicted trajectories associated with the respective target location.

4 . The method of claim 1 , wherein for each target location, generating the first trajectory model includes:

executing a generative behaviour model for the respective target location, the generative behaviour model having been trained to generate trajectories based on examples of real-world driving behaviour.

5 . The method of claim 1 , wherein each first trajectory model is determined based on one or more ego vehicle parameters to model a response of the external actor to ego vehicle behaviour.

6 . The method of claim 1 , further comprising generating a second trajectory model for the respective target location based on the observed trace of the external actor between the first time and the second time,

wherein the determining the likelihood is based on the second trajectory model.

7 . The method of claim 6 , wherein generating the second trajectory model for the respective target location includes:

determining a sequence of possible manoeuvres the external actor to reach the respective target location, wherein the possible manoeuvers are associated, respectively, with partial trajectory models;

wherein the second trajectory model for the respective target location is generated based on the partial trajectory models associated with the possible manoeuvres.

8 . The method of claim 7 , wherein each partial trajectory model of the partial trajectory models comprises one or more target motion values, and wherein generating the second trajectory model for the respective target location comprises applying motion smoothing to the one or more target motion values of the partial trajectory models to determine one or more motion values of a portion of the second trajectory model.

9 . The method of claim 6 , wherein determining the likelihood that the external actor is attempting to reach the respective target location includes:

applying a cost function to both the first trajectory model for the respective target location and the second trajectory model for the respective target location, to determine a first cost associated with the first trajectory model and a second cost associated with the second trajectory model; and

comparing the first cost and the second cost.

10 . The method of claim 1 , further comprising determining at least one predicted trajectory for the external actor after the third time for at least one target location of the set of available target locations; and predicting a probability of the at least one predicted trajectory based on the likelihood that the external actor is attempting to reach the at least one target location.

11 . A computer system for probabilistically predicting a target location of an external actor, the computer system comprising:

one or more processors; and

memory coupled to the one or more processors, the memory embodying computer-readable instructions, which, when executed on the one or more processors, cause the one or more processors to perform operations including:

receiving sensor inputs for detecting and tracking the external actor;

applying object tracking to the sensor inputs, in order to track the external actor, and thereby determine an observed trace of the external actor over a time interval from a first time to a third time;

determining based on a map a set of available target locations for the external actor;

for each target location of the set of available target locations, generating a respective first trajectory model over said time interval from the first time to the third time, based on a location of the external actor at the first time, and the respective target location on the map, wherein each first trajectory model includes one or more first trajectories from the location of the external actor at the first time to the respective target location;

for each target location, determining a likelihood that the external actor is attempting to reach the respective target location based on (1) the first trajectory model for the respective target location and (2) an observed trace of the external actor between the first time and a second time, wherein the second time is between the first time and the third time; and

generating one or more control signals based on the determined likelihood, and providing the one or more control signals to a drive mechanism of an autonomous vehicle.

12 . The computer system of claim 11 , wherein;

the computer system is an autonomous vehicle computer system;

the receiving the sensor inputs, the applying the object tracking, the determining the set of available target locations, the generating the generating the respective first trajectory model for each target location, and the determining the likelihood are performed by a prediction component; and

the operations further include making, by a planner, autonomous driving decisions using outputs of the prediction component.

13 . The computer system of claim 12 , wherein the computer system is configured to be installed on-board the autonomous vehicle, wherein the planner is configured to be coupled to the drive mechanism of the autonomous vehicle, the planner being configured to generate the one or more control signals and provide the one or more control signals to the drive mechanism.

14 . A non-transitory computer readable medium embodying computer program instructions, the computer program instructions configured so as, when executed on one or more hardware processors, to implement operations comprising:

receiving, at a computer, sensor inputs for detecting and tracking an external actor;

applying object tracking to the sensor inputs, in order to track the external actor, and thereby determine an observed trace of the external actor over a time interval from a first time to a third time;

determining based on a map a set of available target locations for the external actor;

for each target location of the set of available target locations, generating a respective first trajectory model over said time interval from the first time to the third time, based on a location of the external actor at the first time, and the respective target location on the map;

for each target location, determining a likelihood that the external actor is attempting to reach the respective target location based on (1) the first trajectory model for the respective target location and (2) an observed trace of the external actor between the first time and a second time, wherein the second time is between the first time and the third time; and

generating one or more control signals based on the determined likelihood, and providing the one or more control signals to a drive mechanism of an autonomous vehicle.

15 . The method of claim 1 , wherein for each target location, generating the first trajectory model for the respective target location includes:

applying a sampling algorithm to sample a space of predicted trajectories defined for the respective target location based on one or more parameters of the respective target location and one or more parameters of the external actor.