Assessing present intentions of an actor perceived by an autonomous vehicle

1. A method of forecasting an intention of an actor in an environment through which an autonomous vehicle is traveling, the method comprising:

by a perception system of an autonomous vehicle:

detecting an actor that is proximate to the autonomous vehicle,

determining a class of the actor, and

detecting an action that the actor is taking;

by a forecasting system of the autonomous vehicle for each cycle of a plurality of cycles:

using the class and the detected action to generate a plurality of candidate intentions of the actor,

evaluating a likelihood of each said candidate intention, and

saving each of the candidate intentions and their likelihoods in a data store;

after any current cycle of the plurality of cycles has completed, analyzing the candidate intentions and their likelihoods for the current cycle and for one or more of the prior cycles to determine an overall probability for each candidate intention of the candidate intentions based on at least a persistence of the candidate intention over a non-interrupted sequence of cycles, where each said cycle represents a time period over which the actor was sensed by the perception system; and

by a motion planning system of the autonomous vehicle;

using the overall probabilities to select at least one candidate intention of the candidate intentions,

forecasting a subsequent future intention that the actor may have after reaching a goal defined by the selected at least one candidate intention,

obtaining an actor trajectory that is consistent with the selected at least one candidate intention and the subsequent future intention, and

using the actor trajectory to influence a selected trajectory for the autonomous vehicle.

2. The method of claim 1 , wherein using the class and the detected action to generate a plurality of candidate intentions of the actor comprises:

accessing a data set of possible goals that are associated with various classes of actors;

selecting possible goals that the data set associates with the detected class of the actor;

determining which of the possible goals in the data set are consistent with the detected action; and

using the determined possible goals as the candidate intentions.

3. The method of claim 2 , wherein determining which of the possible goals in the data set are consistent with the detected action comprises:

determining whether the detected action satisfies one or more rules of each of the possible goals; or

processing the detected actions and possible goals in a machine learning model that has been trained on a data set of labeled actions and goals.

4. The method of claim 2 , wherein determining the overall probabilities for the candidate intentions comprises:

identifying a first goal of the candidate intentions and a second goal of the candidate intentions, wherein the first goal conflicts with the second goal; and

determining that the detected actions of the actor over a plurality of cycles were consistent with the first goal, and in response reducing the overall probability of the candidate intention that is the second goal.

5. The method of claim 1 , wherein evaluating the likelihood of each candidate intention comprises tracking each candidate intention with a unique probabilistic model per candidate intention, in which [false, true] states of each probabilistic model represent whether the candidate intention corresponds to detected action.

6. The method of claim 1 , further comprising, after an additional group of cycles have been completed, wherein the additional group comprises at least some of the cycles considered when determining the overall probabilities plus one or more additional cycles:

analyzing the candidate intentions and their likelihoods for the additional group of cycles to refine the overall probabilities for each of the candidate intentions.

7. The method of claim 1 , further comprising:

assigning a probability to the forecasted subsequent future intention;

wherein the actor trajectory is obtained using the forecasted future intention when the probability exceeds a threshold.

8. The method of claim 7 , further comprising, before forecasting the subsequent future intention of the actor:

eliminating candidate intentions having likelihoods that are below a threshold, so that forecasting the future subsequent intention of the actor is only performed for the at least one candidate intention which has a relatively high likelihood.

9. The method of claim 1 , wherein determining the overall probabilities for each of the candidate intentions comprises:

assigning a relatively higher likelihood to any candidate intention that persisted over a non-interrupted sequence of cycles; and

assigning a relatively lower likelihood to any candidate intention that did not persist over a non-interrupted sequence of cycles.

10. The method of claim 1 , wherein determining the overall probabilities for each of the candidate intentions comprises:

evaluating each of the candidate intentions against a kinematic state of the actor;

assigning a relatively higher likelihood to any candidate intention that is consistent with the kinematic state of the actor; and

assigning a relatively lower likelihood to any candidate intention that is not consistent with the kinematic state of the actor.

11. The method of claim 1 , wherein determining the overall probabilities for the candidate intentions comprises:

receiving, from a plurality of sensors of the perception system, environmental data relating to the environment through which the autonomous vehicle is traveling;

applying the environmental data to a random forest classifier to classify an environmental condition of the environment;

assigning a relatively higher likelihood to any candidate intention that is consistent with the environmental condition; and

assigning a relatively lower likelihood to any candidate intention that is not consistent with the environmental condition.

12. An autonomous vehicle, comprising:

a perception system comprising:

one or more sensors for capturing perception data about actors that are proximate to the autonomous vehicle,

a processor, and

a memory with programming instructions configured to instruct the processor of the perception system to process the perception data to:

detect and classify an actor that is proximate to the autonomous vehicle, and

detect an action that the actor is taking;

a forecasting system comprising:

a processor, and

a memory with programming instructions configured to cause the processor of the forecasting system to:

for each cycle of a plurality of cycles:

use the class and the detected action to generate a plurality of candidate intentions of the actor;

evaluate a likelihood of each candidate intention; and

save each of the candidate intentions and their likelihoods in a data store,

after any current cycle of the plurality of cycles has completed, analyze the candidate intentions and their likelihoods for the current cycle and for one or more of the prior cycles to determine an overall probability for each intention of the candidate intentions based on at least a persistence of the intention over a non-interrupted sequence of cycles, where each said cycle represents a time period over which the actor was sensed by the perception system; and

a motion planning system comprising:

a processor, and

a memory with programming instructions configured to cause the processor of the motion planning system to:

use the overall probabilities to select at least one candidate intention of the candidate intentions,

forecast a subsequent future intention that the actor may have after reaching a goal defined by the selected at least one candidate intention,

obtain an actor trajectory that is consistent with the selected at least one candidate intention and the subsequent future intention, and

use the actor trajectory to influence a selected trajectory for the autonomous vehicle.

13. The autonomous vehicle of claim 12 , wherein the instructions to use the class and the detected action to generate a plurality of candidate intentions of the actor comprise instructions to:

access a data set of possible goals that are associated with various classes of actors;

select possible goals that the data set associates with the detected class of the actor;

determine which of the possible goals in the data set are consistent with the detected action; and

use the determined possible goals as the candidate intentions.

14. The autonomous vehicle of claim 13 , wherein the instructions to determine which of the possible goals in the data set are consistent with the detected action comprise instructions to:

determine whether the detected action satisfies one or more rules of each of the possible goals; or

process the detected actions and possible goals in a machine learning model that has been trained on a data set of labeled actions and goals.

15. The autonomous vehicle of claim 12 , wherein the instructions to evaluate the likelihood of each candidate intention comprise instructions to track each candidate intention with a unique probabilistic model per candidate intention, in which [false, true] states of each probabilistic model represent whether the candidate intention corresponds to detected action.

16. The autonomous vehicle of claim 12 , further comprising additional programming instructions configured to instruct the processor of the forecasting system to, after an additional group of cycles have been completed, wherein the additional group comprises at least some of the cycles considered when determining the overall probabilities plus one or more additional cycles:

analyze the candidate intentions and their likelihoods for the additional group of cycles to refine the overall probabilities for each of the candidate intentions.

17. The autonomous vehicle of claim 12 , further comprising additional programming instructions that are configured to cause the processor of the forecasting system to, for at least some of the candidate intentions, assign a probability to the forecasted subsequent future intention;

wherein the actor trajectory is obtained using the forecasted subsequent future intention when the probability exceeds a threshold.

18. The autonomous vehicle of claim 12 , further comprising additional programming instructions that are configured to cause the processor of the forecasting system to, before forecasting the subsequent future intention of the actor:

eliminate candidate intentions having likelihoods that are below a threshold, so that forecasting the subsequent future intention of the actor is only performed for the at least one candidate intention which has a relatively high likelihood.

19. The autonomous vehicle of claim 12 , wherein the instructions to determine the overall probabilities for each of the candidate intentions comprise instructions to:

assign a relatively higher likelihood to any candidate intention that persisted over a non-interrupted sequence of cycles; and

assign a relatively lower likelihood to any candidate intention that did not persist over a non-interrupted sequence of cycles.

20. The autonomous vehicle of claim 12 , wherein the instructions to determine the overall probabilities for each of the candidate intentions comprise instructions to:

evaluate each of the candidate intentions against a kinematic state of the actor;

assign a relatively higher likelihood to any candidate intention that is consistent with the kinematic state of the actor; and

assign a relatively lower likelihood to any candidate intention that is not consistent with the kinematic state of the actor.

21. The autonomous vehicle of claim 12 , wherein the instructions to determine the overall probabilities for each of the candidate intentions comprise instructions to:

receive, from one or more of the sensors of the perception system, environmental data relating to the environment through which the autonomous vehicle is traveling;

apply the environmental data to a random forest classifier to classify an environmental condition of the environment;

assign a relatively higher likelihood to any candidate intention that is consistent with the environmental condition; and

assign a relatively lower likelihood to any candidate intention that is not consistent with the environmental condition.

22. The autonomous vehicle of claim 12 , wherein the instructions to determine the overall probabilities for each of the candidate intentions comprise instructions to:

identify a first goal of the candidate intentions and a second goal of the candidate intentions, wherein the first goal conflicts with the second goal; and

determine that the detected actions of the actor over a plurality of cycles were consistent with the first goal, and in response reduce the overall probability of the candidate intention that is the second goal.

23. A non-transitory computer-readable medium that stores instructions that are configured to, when executed by at least one computing device, cause the at least one computing device to perform operations comprising:

detecting an actor that is proximate to the autonomous vehicle,

determining a class of the actor, and

detecting an action that the actor is taking;

for each cycle of a plurality of cycles:

using the class and the detected action to generate a plurality of candidate intentions of the actor,

evaluating a likelihood of each candidate intention, and

saving each of the candidate intentions and their likelihoods in a data store;

after any current cycle of the plurality of cycles has completed, analyzing the candidate intentions and their likelihoods for the current cycle and for one or more of the prior cycles to determine an overall probability for each intention of the candidate intentions based on at least a persistence of the intention over a non-interrupted sequence of cycles, where each said cycle represents a time period over which the actor was sensed by the perception system; and

using the overall probabilities to select at least one candidate intention of the candidate intentions,

forecasting a subsequent future intention that the actor may have after reaching a goal defined by the selected at least one candidate intention,

obtaining an actor trajectory that is consistent with the selected at least one candidate intention and the subsequent future intention, and

using the actor trajectory to influence a selected trajectory for the autonomous vehicle.