Methods and systems for long-term trajectory prediction by extending a prediction horizon
Systems and methods for planning a trajectory for a vehicle are disclosed. The methods include generating first predicted states of an object in an environment of the vehicle at one or more first time steps during a first prediction horizon using a machine learning model and sensor data about the environment. The first predicted states are used as an input to a physical model to generate second predicted states of the object at one or more second time steps during a second prediction horizon where the one or more second time steps are after the one or more first time steps. A trajectory for the vehicle is generated using the first predicted states or the second predicted states of the object.
1. A method for planning a trajectory for a vehicle, the method comprising, by a processor:
generating, using a machine learning model and sensor data about an environment, first predicted states of an object at one or more first time steps during a first prediction horizon;
using the first predicted states of the object as an input to a physical model to generate second predicted states of the object at one or more second time steps during a second prediction horizon, the one or more second time steps being after the one or more first time steps; generating, using the first predicted states or the second predicted states of the object, the trajectory for the vehicle; and
controlling the vehicle to traverse the environment using the trajectory generated.
2. The method of claim 1 , further comprising:
receiving the sensor data; and
determining, based on the sensor data, one or more features of the object, the one or more features comprising past or current observed states of the object.
3. The method of claim 1 , wherein the first predicted states of the object comprise at least a predicted path of the object during the first prediction horizon.
4. The method of claim 1 , further comprising selecting, based on one or more features of the object, the physical model from a plurality of physical models.
5. The method of claim 1 , wherein the first prediction horizon is about 4-12 seconds.
6. The method of claim 1 , wherein the second prediction horizon is about 8-15 seconds.
7. The method of claim 1 , further comprising determining a reference path for the object.
8. The method of claim 1 , further comprising:
selectively disabling the machine learning model; and
generating, using the physical model and sensor data about an environment, a third predicted state of the object.
9. The method of claim 1 , wherein the machine learning model is a deep learning model.
10. A system for planning a trajectory for a vehicle, the system comprising:
at least one processor;
programming instructions stored in a memory and configured to cause the processor to:
generate, using a machine learning model and sensor data about an environment, first predicted states of an object at one or more first time steps during a first prediction horizon;
use the first predicted states of the object as an input to a physical model to generate second predicted states of the object at one or more second time steps during a second prediction horizon, the one or more second time steps being after the one or more first time steps;
generate, using the first predicted states or the second predicted states of the object, the trajectory for the vehicle; and
control the vehicle to traverse the environment using the trajectory generated.
11. The system of claim 10 , further comprising programming instructions configured to cause the processor to:
receive the sensor data; and
determine, based on the sensor data, one or more features of the object, the one or more features comprising past or current observed states of the object.
12. The system of claim 10 , wherein the first predicted states of the object comprise at least a predicted path of the object during the first prediction horizon.
13. The system of claim 10 , further comprising programming instructions configured to cause the processor to select, based on one or more features of the object, the physical model from a plurality of physical models.
14. The system of claim 10 , wherein the first prediction horizon is about 4-12 seconds.
15. The system of claim 10 , wherein the second prediction horizon is about 8-15 seconds.
16. The system of claim 10 , further comprising programming instructions configured to cause the processor to determine a reference path for the object.
17. The system of claim 10 , further comprising programming instructions configured to cause the processor to:
selectively disable the machine learning model; and
generate, using the physical model and sensor data about an environment, a third predicted state of the object.
18. The system of claim 10 , wherein the machine learning model is a deep learning model.
19. A computer program product comprising a non-transitory computer-readable medium that stores instructions that, when executed by a computing device, will cause the computing device to perform operations comprising:
generating, using a machine learning model and sensor data about an environment, first predicted states of an object at one or more first time steps during a first prediction horizon;
using the first predicted states of the object as an input to a physical model to generate second predicted states of the object at one or more second time steps during a second prediction horizon, the one or more second time steps being after the one or more first time steps;
generating, using the first predicted states or the second predicted states of the object, a trajectory for a vehicle; and
controlling the vehicle using the trajectory generated.
20. The computer program product of claim 19 , further comprising instructions to perform operations comprising:
receiving the sensor data; and
determining, based on the sensor data, one or more features of the object, the one or more features comprising past or current observed states of the object.