Reinforcement learning with scene decomposition for navigating complex environments
Systems and methods for providing navigation to a vehicle may include receiving observation data from one or more sensors of the vehicle, generating projection data corresponding to the one or more traffic participants based on the observation data for each time step within a time period, and predicting interactions between the vehicle, the one or more traffic participants, and the one or more obstacles, based on the projection data of the one or more traffic participants. The systems and methods may further include determining a set of actions by the vehicle corresponding to a probability of the vehicle safely arriving at a target location based on the predicted interactions, and selecting one or more actions from the set of actions and provide the one or more actions to a navigation system of the vehicle, wherein the navigation system uses the navigation data to provide navigation instructions to the vehicle.
1. A method of providing navigation for a vehicle performed by a computing device of a navigation system, comprising:
receiving, from one or more sensors mounted to the vehicle, observation data including state information of the vehicle, one or more traffic participants detected by the one or more sensors, and one or more obstacles detected by the one or more sensors;
predicting one or more potential traffic participants that are unobserved by the one or more sensors due to the one or more obstacles;
generating projection data corresponding to the one or more traffic participants based on the observation data for each time step within a time period, and corresponding to the one or more potential traffic participants that are unobserved by the one or more sensors due to the one or more obstacles;
predicting potential interactions between the vehicle, the one or more traffic participants, and the one or more potential traffic participants that are unobserved by the one or more sensors due to the one or more obstacles, based on the projection data;
determining a set of actions by the vehicle corresponding to a probability of the vehicle safely arriving at a target location based on the potential interactions;
selecting one or more actions from the set of actions; and
sending, to a component of the navigation system, data corresponding to the one or more actions, wherein the navigation system provides navigation instructions to the vehicle in response to the data corresponding to the one or more actions.
2. The method of claim 1 , wherein the selecting of the one or more actions comprises:
determining a set of timed actions by the vehicle corresponding a time of the vehicle to arrive at the target location; and
selecting the one or more actions further based on the set of timed actions.
3. The method of claim 2 , wherein the determining of the set of timed actions is performed by a trained deep Q-learning network (DQN).
4. The method of claim 1 , wherein the selecting of the one or more actions from the set of actions is based on the one or more actions being greater than a threshold.
5. The method of claim 1 , further comprising:
receiving historical observation data, wherein the generating of the projection data is further based on the historical observation data.
6. The method of claim 1 , wherein one or more of the generating of the projection data or the predicting of the potential interactions are performed by a plurality of recurrent neural networks (RNNs) trained on different portions of a dataset.
7. A vehicle computing device for a navigation system of a vehicle, comprising:
a sensor system including one or more sensors mounted to the vehicle and configured to collect observation data;
a memory storing executable instructions; and
a processor coupled with the memory and configured to:
receive, from the sensor system, observation data including state information of the vehicle, one or more traffic participants detected by the one or more sensors, and one or more obstacles detected by the one or more sensors;
predict one or more potential traffic participants that are unobserved by the one or more sensors due to the one or more obstacles;
generate projection data corresponding to the one or more traffic participants based on the observation data for each time step within a time period, and corresponding to the one or more potential traffic participants that are unobserved by the one or more sensors due to the one or more obstacles;
predict potential interactions between the vehicle, the one or more traffic participants, and the one or more potential traffic participants that are unobserved by the one or more sensors due to the one or more obstacles, based on the projection data of the one or more traffic participants;
determine a set of actions by the vehicle corresponding to a probability of the vehicle safely arriving at a target location based on the potential interactions;
select one or more actions from the set of actions; and
autonomously controlling the vehicle based on the one or more actions.
8. The vehicle computing device of claim 7 , wherein the processor is further configured to:
determine a set of timed actions by the vehicle corresponding a time of the vehicle to arrive at the target location; and
selected the one or more actions further based on the set of timed actions.
9. The vehicle computing device of claim 8 , wherein the processor includes a trained deep Q-learning network (DQN) to determine the set of timed actions.
10. The vehicle computing device of claim 7 , wherein the one or more actions is selected based on the one or more actions being greater than a threshold.
11. The vehicle computing device of claim 7 , wherein the processor is further configured to:
receive historical observation data, wherein the projection data is generated further based on the historical observation data.
12. The vehicle computing device of claim 7 , wherein the processor includes a plurality of recurrent neural networks (RNNs) trained on different portions of a dataset to generate the projection data or predict the potential interactions.
13. A non-transitory computer-readable medium storing computer executable code for a computing device of a navigation system of a vehicle, comprising code to:
receive, from one or more sensors mounted to the vehicle, observation data including state information of the vehicle, one or more traffic participants detected by the one or more sensors, and one or more obstacles detected by the one or more sensors;
predict one or more potential traffic participants that are unobserved by the one or more sensors due to the one or more obstacles;
generate projection data corresponding to the one or more traffic participants based on the observation data for each time step within a time period, and corresponding to the one or more potential traffic participants that are unobserved by the one or more sensors due to the one or more obstacles;
predict potential interactions between the vehicle, the one or more traffic participants, and the one or more potential traffic participants that are unobserved by the one or more sensors due to the one or more obstacles, based on the projection data;
determine a set of actions by the vehicle corresponding to a probability of the vehicle safely arriving at a target location based on the potential interactions;
select one or more actions from the set of actions; and
send, to a component of the navigation system, data corresponding to the one or more actions, wherein the navigation system provides navigation instructions to the vehicle in response to the data corresponding to the one or more actions.
14. The non-transitory computer-readable medium of claim 13 , further comprising code to:
determine a set of timed actions by the vehicle corresponding a time of the vehicle to arrive at the target location; and
select the one or more actions further based on the set of timed actions.
15. The non-transitory computer-readable medium of claim 13 , further comprising code to:
select the one or more actions from the set of actions based on the one or more actions being greater than a threshold.
16. The non-transitory computer-readable medium of claim 13 , further comprising code to:
receive historical observation data; and
generate the projection data further based on the historical observation data.
17. The non-transitory computer-readable medium of claim 13 , further comprising code to:
perform one or more of generating of the projection data or predicting of the potential interactions by a plurality of recurrent neural networks (RNNs) trained on different portions of a dataset.