Systems and methods for training probabilistic object motion prediction models using non-differentiable prior knowledge
The present disclosure provides systems and methods for training probabilistic object motion prediction models using non-differentiable representations of prior knowledge. As one example, object motion prediction models can be used by autonomous vehicles to probabilistically predict the future location(s) of observed objects (e.g., other vehicles, bicyclists, pedestrians, etc.). For example, such models can output a probability distribution that provides a distribution of probabilities for the future location(s) of each object at one or more future times. Aspects of the present disclosure enable these models to be trained using non-differentiable prior knowledge about motion of objects within the autonomous vehicle's environment such as, for example, prior knowledge about lane or road geometry or topology and/or traffic information such as current traffic control states (e.g., traffic light status).
1. A computer-implemented method, comprising:
obtaining, by a computing system comprising one or more computing devices, sensor data descriptive of an environment comprising an object;
processing, by the computing system, the sensor data with a machine-learned object motion prediction model to obtain a predicted location probability distribution for a future location of the object at one or more future times;
sampling, by the computing system, a plurality of sample trajectories from the predicted location probability distribution;
evaluating, by the computing system and for each sample trajectory, a non-differentiable prior knowledge reward function that encodes prior knowledge about motion of the object to obtain a respective reward value for the sample trajectory;
determining, by the computing system, an approximate gradient of an expected loss based at least in part on the reward values respectively obtained for the plurality of sample trajectories;
modifying, by the computing system, one or more values of one or more parameters of the machine-learned object motion prediction model based at least in part on the approximate gradient of the expected loss; and
controlling, by the computing system, motion of an autonomous vehicle based at least in part on the modified one or more values of one or more parameters of the machine-learned object motion prediction model.
2. The computer-implemented method of claim 1 , wherein:
the sensor data descriptive of the environment comprises real world sensor data collected by sensors associated with the autonomous vehicle while the autonomous vehicle is operated; and
the non-differentiable prior knowledge reward function compares the sample trajectory with ground truth data associated with the object.
3. The computer-implemented method of claim 1 , wherein the prior knowledge about motion of the object comprises prior knowledge about lane geometry, road topology, or traffic rules within the environment.
4. The computer-implemented method of claim 1 , wherein evaluating, by the computing system and for each sample trajectory, the non-differentiable prior knowledge reward function comprises:
determining, by the computing system, a reachable area for the object, wherein the reachable area is defined by a set of one or more reachable lanes that are reachable from a current location of the object while observing traffic rules; and
evaluating, by the computing system and for each sample trajectory, the non-differentiable prior knowledge reward function based on the reachable area.
5. The computer-implemented method of claim 4 , wherein:
the non-differentiable prior knowledge reward function returns a positive reward when the sample trajectory stays within the reachable area; and
the non-differentiable prior knowledge reward function returns a negative reward when the sample trajectory exits the reachable area.
6. The computer-implemented method of claim 4 , wherein:
the non-differentiable prior knowledge reward function returns a positive reward when the sample trajectory stays within the reachable area and a ground truth trajectory associated with the object stays within a ground truth reachable area; and
the non-differentiable prior knowledge reward function returns a negative reward when the sample trajectory exits the reachable area and the ground truth trajectory associated with the object stays within the ground truth reachable area.
7. The computer-implemented method of claim 4 , wherein the traffic rules comprise lane infraction rules that prohibit crossing a solid line or entering a lane having an opposite traffic flow direction.
8. The computer-implemented method of claim 4 , wherein the traffic rules comprise adhering to a current traffic control state provided by a traffic control device.
9. The computer-implemented method of claim 1 , wherein evaluating, by the computing system and for each sample trajectory, the non-differentiable prior knowledge reward function comprises determining, by the computing system and for each sample trajectory, whether the sample trajectory intersects with a route associated with an ego vehicle, wherein the reward value is a function of whether the sample trajectory intersects with the route.
10. The computer-implemented method of claim 9 , wherein:
the non-differentiable prior knowledge reward function returns a true positive reward value when the sample trajectory intersects the route and a ground truth trajectory associated with the object intersects the route;
the non-differentiable prior knowledge reward function returns a false positive reward value when the sample trajectory intersects the route and the ground truth trajectory associated with the object does not intersect the route;
the non-differentiable prior knowledge reward function returns a true negative reward value when the sample trajectory does not intersect the route and the ground truth trajectory associated with the object does not intersect the route; and
the non-differentiable prior knowledge reward function returns a false positive reward value when the sample trajectory does not intersect the route and the ground truth trajectory associated with the object intersects the route.
11. The computer-implemented method of claim 9 , wherein the route associated with the ego vehicle comprises a current motion plan for the ego vehicle.
12. The computer-implemented method of claim 1 , wherein determining, by the computing system, the approximate gradient of the expected loss based at least in part on the reward values respectively obtained for the plurality of sample trajectories comprises performing, by the computing system, a REINFORCE gradient estimation technique.
13. The computer-implemented method of claim 1 , wherein the one or more future times comprise a plurality of future times.
14. The computer-implemented method of claim 1 , wherein the machine-learned object motion prediction model comprises a spatially-aware graph neural network combined with a multi-layer perceptron parameterized as a mixture of Gaussians.
15. The computer-implemented method of claim 1 , wherein modifying, by the computing system, the one or more values of the one or more parameters of the machine-learned object motion prediction model based at least in part on the approximate gradient of the expected loss comprises modifying, by the computing system, the one or more values of the one or more parameters of the machine-learned object motion prediction model based at least in part on a heuristically chosen closest matching mode.
16. A computing system, comprising:
one or more processors; and
one or more non-transitory computer-readable mediums that collectively store instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising:
obtaining, by the computing system, sensor data descriptive of an environment comprising an object;
processing, by the computing system, the sensor data with a machine-learned object motion prediction model to obtain a predicted location probability distribution for a future location of the object at one or more future times;
sampling, by the computing system, a plurality of sample trajectories from the predicted location probability distribution;
evaluating, by the computing system and for each sample trajectory, a non-differentiable prior knowledge reward function that encodes prior knowledge about motion of the object to obtain a respective reward value for the sample trajectory;
determining, by the computing system, an approximate gradient of an expected loss based at least in part on the reward values respectively obtained for the plurality of sample trajectories;
modifying, by the computing system, one or more values of one or more parameters of the machine-learned object motion prediction model based at least in part on the approximate gradient of the expected loss; and
controlling, by the computing system, motion of an autonomous vehicle based at least in part on the modified one or more values of one or more parameters of the machine-learned object motion prediction model.
17. The computing system of claim 16 , wherein:
the sensor data descriptive of the environment comprises real world sensor data collected by sensors associated with the autonomous vehicle while the autonomous vehicle is operated; and
the non-differentiable prior knowledge reward function compares the sample trajectory with ground truth data associated with the object.
18. The computing system of claim 16 , wherein the prior knowledge about motion of the object comprises prior knowledge about lane geometry or topology within the environment.
19. The computing system of claim 18 , wherein evaluating, by the computing system and for each sample trajectory, the non-differentiable prior knowledge reward function comprises:
determining, by the computing system, a reachable area for the object, wherein the reachable area is defined by a set of one or more reachable lanes that are reachable from a current location of the object while observing traffic rules; and
evaluating, by the computing system and for each sample trajectory, the non-differentiable prior knowledge reward function based on the reachable area.
20. An autonomous vehicle computing system, comprising:
one or more processors; and
one or more non-transitory computer-readable mediums that collectively store:
a machine-learned object motion prediction model configured to process sensor data descriptive of an environment comprising an object to generate a predicted location probability distribution for a future location of the object at one or more future times, wherein the machine-learned object motion prediction model has been trained by performing a REINFORCE gradient estimation technique to determine an approximate gradient of an expected loss that is a function of a non-differentiable prior knowledge reward function that encodes prior knowledge about motion of the object; and
instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising:
obtaining the sensor data for the environment comprising the object;
processing the sensor data with the machine-learned object motion prediction model to generate the predicted location probability distribution for the future location of the object at one or more future times; and
controlling motion of an autonomous vehicle based at least in part on the predicted location probability distribution.