Method and system for data-driven and modular decision making and trajectory generation of an autonomous agent

1. A method for decision making and trajectory generation of an autonomous agent, the method comprising:

receiving a set of inputs, wherein the set of inputs comprises:

a context of the autonomous agent, wherein the context of the autonomous agent is determined based on a labeled map comprising a set of predetermined context assignments;

with a set of one or more computing subsystems:

selecting a first learned model from a first set of multiple learned models based on the context;

determining an action for the autonomous agent based on the first learned model;

selecting a second learned model from a second set of multiple learned models based on the action;

determining a trajectory for the autonomous agent based on the second learned model;

validating the trajectory; and

operating the autonomous agent based on the trajectory.

2. The method of claim 1 , wherein each of the first set of multiple learned models is mapped in a 1:1 mapping to a context of the autonomous agent.

3. The method of claim 1 , wherein the set of inputs further comprises a pose of the autonomous agent, wherein the context is further determined based on the pose.

4. The method of claim 3 , wherein the set of predetermined context assignments is determined based on a fixed route assignment of the autonomous agent, wherein the trajectory is determined at least in part based on the fixed route assignment.

5. The method of claim 1 , wherein each of the first and second learned models is a deep Q-learning network trained with an inverse reinforcement learning algorithm.

6. The method of claim 1 , wherein the second learned model is selected based on a mapping from the action to the second learned model.

7. The method of claim 6 , wherein the mapping is a 1:1 mapping.

8. The method of claim 1 , wherein the trajectory is validated based on checking for at least one of:

an uncertainty associated with the trajectory having a value below a predetermined threshold; and

a satisfaction of a programmed set of rules, wherein checking for satisfaction of the programmed set of rules comprises checking for at least one of: an absence of a collision of the trajectory with a static object, an absence of a collision of the trajectory with a dynamic object, satisfaction of the trajectory with a set of traffic laws, and satisfaction of the trajectory with road infrastructure information.

9. The method of claim 1 , wherein a set of parameter values associated with the action is determined based on the context.

10. The method of claim 9 , wherein the set of parameter values comprises at least one of a creep distance parameter and a stopping distance parameter.

11. A method for decision making and trajectory generation of an autonomous agent, the method comprising:

receiving a set of inputs, wherein the set of inputs comprises:

a pose of the autonomous agent;

routing information associated with a fixed route assignment of the autonomous agent; and

a labeled map;

static object information; and

dynamic object information;

selecting a first learned model from a first set of multiple learned models based on the set of inputs;

selecting a second learned model from a second set of multiple learned models based on an output of the first learned model;

determining a trajectory for the autonomous agent based on the second learned model;

validating the trajectory with a set of rules; and

operating the autonomous agent based on the trajectory.

12. The method of claim 11 , further comprising determining a context of the autonomous agent based on the set of inputs, wherein the first learned model is selected additionally based on the context.

13. The method of claim 12 , wherein the context is determined based on the labeled map, wherein the labeled map comprises a set of predetermined context assignments.

14. The method of claim 13 , wherein the first learned model and the context are associated with each other in a 1:1 mapping.

15. The method of claim 11 , further comprising determining a localized environmental representation based on the set of inputs, wherein the localized environmental representation is received at the second learned model.

16. The method of claim 11 , wherein the set of rules comprises a first subset of rules and a second subset of rules, wherein the method further comprises:

checking for the first subset of rules, wherein checking for the first subset of rules comprises:

checking for an absence of a collision of the autonomous agent with one or more static objects based on the static object information and the trajectory;

checking for an absence of a collision of the autonomous agent with one or more dynamic objects based on the dynamic object information and the trajectory; and

checking for the second subset of rules, wherein checking for the second subset of rules comprises:

checking for satisfaction of the trajectory with at least one of traffic laws and the road infrastructure information.

17. The method of claim 11 , further comprising defining a safety tunnel of the autonomous agent based on the action, wherein the safety tunnel prescribes at least one of:

a predetermined and fixed distance relative to the autonomous agent; and

a distance to a stopping point of the autonomous agent.

18. The method of claim 17 , further comprising determining a localized environmental representation based on the set of inputs and the safety tunnel, wherein the localized environmental representation is an input to the selected second learned model.

19. The method of claim 18 , wherein determining the localized environmental representation comprises removing static object information and dynamic object information outside of the safety tunnel.

20. The method of claim 19 , wherein determining the localized environmental representation further comprises determining a latent space representation.