Systems and methods for motion forecasting and planning for autonomous vehicles
Systems and methods are disclosed for motion forecasting and planning for autonomous vehicles. For example, a plurality of future traffic scenarios are determined by modeling a joint distribution of actor trajectories for a plurality of actors, as opposed to an approach that models actors individually. As another example, a diversity objective is evaluated that rewards sampling of the future traffic scenarios that require distinct reactions from the autonomous vehicle. An estimated probability for the plurality of future traffic scenarios can be determined and used to generate a contingency plan for motion of the autonomous vehicle. The contingency plan can include at least one initial short-term trajectory intended for immediate action of the AV and a plurality of subsequent long-term trajectories associated with the plurality of future traffic scenarios.
1. A computer-implemented method for motion forecasting and planning, the method comprising:
(a) determining a plurality of actors within an environment of an autonomous vehicle from sensor data descriptive of the environment;
(b) determining a plurality of future motion scenarios based on the sensor data by modeling a joint distribution of predicted actor trajectories as a latent variable model for the plurality of actors, wherein determining the plurality of future motion scenarios comprises: (i) evaluating a diversity objective that rewards sampling of the plurality of future motion scenarios that require distinct reactions from the autonomous vehicle, and (ii) obtaining a particular number of latent samples from the latent variable model based on the diversity objective;
(c) determining an estimated probability for the plurality of future motion scenarios; and
(d) generating a contingency plan for motion of the autonomous vehicle, wherein the contingency plan comprises at least one initial short-term trajectory and a plurality of subsequent long-term trajectories associated with the plurality of future motion scenarios, and wherein the contingency plan is generated based on the plurality of future motion scenarios and the estimated probability for the plurality of future motion scenarios.
2. The computer-implemented method of claim 1 , wherein (a) comprises processing features from the sensor data and corresponding map data with a first machine-learned model to generate one or more object detections corresponding to the plurality of actors.
3. The computer-implemented method of claim 2 , wherein (a) comprises processing the one or more object detections with a second machine-learned model to generate a respective feature vector defining a local context for one or more of the plurality of actors.
4. The computer-implemented method of claim 3 , wherein the first machine-learned model, the second machine-learned model, an encoder, and a prediction decoder are jointly trained for object detection and motion forecasting.
5. The computer-implemented method of claim 1 , wherein (b) comprises mapping a shared noise across a joint set of latent variables that represent the joint distribution of the predicted actor trajectories for the plurality of actors to determine the plurality of future motion scenarios.
6. The computer-implemented method of claim 5 , wherein (b) comprises employing a graph neural network (GNN) for the mapping of the shared noise across the joint set of latent variables.
7. The computer-implemented method of claim 1 , wherein (b) comprises evaluating an energy function comprising one or more energy terms configured to promote diversity among the plurality of future motion scenarios.
8. The computer-implemented method of claim 6 , wherein (c) comprises employing the GNN to output a score corresponding to the estimated probability for the plurality of future motion scenarios.
9. The computer-implemented method of claim 1 , wherein (d) comprises optimizing a planner cost function comprising a linear combination of subcosts that encode different aspects of driving, the different aspects of driving comprising two or more of comfort, motion rules, or route.
10. The computer-implemented method of claim 1 , wherein (d) comprises:
(i) generating a plurality of paths;
(ii) determining a set of initial short-term trajectories by sampling a first set of velocity profiles for the plurality of paths; and
(iii) determining a set of subsequent long-term trajectories by sampling a second set of velocity profiles that are conditioned on an end state of the set of initial short-term trajectories.
11. An autonomous vehicle (AV) control system, comprising:
one or more processors; and
one or more non-transitory computer-readable media that store instructions for execution by the one or more processors to cause the AV control system to perform operations, the operations comprising:
(a) determining a plurality of actors within an environment of an autonomous vehicle, wherein the plurality of actors are determined from sensor data descriptive of the environment;
(b) determining a plurality of future traffic scenarios based on the sensor data, wherein the plurality of future traffic scenarios are determined by modeling a joint distribution of actor trajectories as a latent variable model for the plurality of actors, wherein determining the plurality of future traffic scenarios comprises: (i) evaluating a diversity objective that rewards sampling of the plurality of future traffic scenarios that require distinct reactions from the autonomous vehicle, and (ii) obtaining a particular number of latent samples from the latent variable model based on the diversity objective;
(c) determining an estimated probability for the plurality of future traffic scenarios;
(d) generating a contingency plan for motion of the autonomous vehicle, wherein the contingency plan comprises at least one initial short-term trajectory and a plurality of subsequent long-term trajectories associated with the plurality of future traffic scenarios, and wherein the contingency plan is generated based on the plurality of future traffic scenarios and the estimated probability for the plurality of future traffic scenarios.
12. The AV control system of claim 11 , wherein (a) comprises employing a first machine-learned model configured to generate multi-class object detections and a second machine-learned model configured to generate respective feature vectors defining a local context for one or more of the plurality of actors.
13. The AV control system of claim 11 , wherein (b) comprises mapping a shared noise across a joint set of latent variables that represent the joint distribution of actor trajectories for the plurality of actors to determine the plurality of future traffic scenarios.
14. The AV control system of claim 13 , wherein (b) comprises employing a graph neural network (GNN) for the mapping of the shared noise across the joint set of latent variables.
15. The AV control system of claim 14 , wherein (c) comprises employing the GNN to output a score corresponding to the estimated probability for the plurality of future traffic scenarios.
16. The AV control system of claim 11 , wherein (d) comprises optimizing a planner cost function comprising a linear combination of subcosts that encode different aspects of driving, the different aspects of driving comprising two or more of comfort, motion rules, or route.
17. The AV control system of claim 11 , wherein (d) comprises:
(i) generating a plurality of paths;
(ii) determining a set of initial short-term trajectories by sampling a first set of velocity profiles for the plurality of paths; and
(iii) determining a set of subsequent long-term trajectories by sampling a second set of velocity profiles that are conditioned on an end state of the set of initial short-term trajectories.
18. An autonomous vehicle, comprising:
one or more processors; and
one or more non-transitory computer-readable media that store instructions for execution by the one or more processors to cause the autonomous vehicle to perform operations, the operations comprising:
(a) determining a plurality of actors within an environment of the autonomous vehicle, wherein the plurality of actors are determined from sensor data descriptive of the environment;
(b) determining a plurality of future traffic scenarios based on the sensor data, wherein the plurality of future traffic scenarios are determined by modeling a joint distribution of actor trajectories as a latent variable model for the plurality of actors, wherein determining the plurality of future traffic scenarios comprises: (i) evaluating a diversity objective that rewards sampling of the plurality of future traffic scenarios that require distinct reactions from the autonomous vehicle, and (ii) obtaining a particular number of latent samples from the latent variable model based on the diversity objective;
(c) determining an estimated probability of the plurality of future traffic scenarios; and
(d) generating a contingency plan for motion of the autonomous vehicle, wherein the contingency plan comprises at least one initial short-term trajectory and a plurality of subsequent long-term trajectories associated with the plurality of future traffic scenarios, and wherein the contingency plan is generated based on the plurality of future traffic scenarios and the estimated probability for the plurality of future traffic scenarios.