End-to-end interpretable motion planner for autonomous vehicles
Systems and methods for generating motion plans including target trajectories for autonomous vehicles are provided. An autonomous vehicle may include or access a machine-learned motion planning model including a backbone network configured to generate a cost volume including data indicative of a cost associated with future locations of the autonomous vehicle. The cost volume can be generated from raw sensor data as part of motion planning for the autonomous vehicle. The backbone network can generate intermediate representations associated with object detections and objection predictions. The motion planning model can include a trajectory generator configured to evaluate one or more potential trajectories for the autonomous vehicle and to select a target trajectory based at least in part on the cost volume generate by the backbone network.
1. An autonomous vehicle, comprising:
one or more processors; and
one or more non-transitory computer-readable media that store:
a machine-learned motion planning model configured to receive sensor data and map data associated with an environment external to an autonomous vehicle and process the sensor data and the map data to generate a target trajectory for the autonomous vehicle, the machine-learned motion planning model configured to:
generate, using the sensor data and the map data, a cost volume that indexes, at each of a plurality of timesteps, cost data associated with a plurality of locations in the environment;
select a target trajectory for the autonomous vehicle from a plurality of potential trajectories based on a comparison of the plurality of potential trajectories using the cost data indexed in the cost volume by, for a respective potential trajectory corresponding to waypoints respectively associated with the plurality of timesteps, using the waypoints and the plurality of timesteps to obtain cost data from the cost volume for determining a respective trajectory score; and
instructions that are executable by the one or more processors to cause the one or more processors to perform operations, the operations comprising:
obtaining the sensor data and the map data; and
processing the sensor data and the map data with the machine-learned motion planning model to receive the target trajectory as an output of the machine-learned motion planning model.
2. The autonomous vehicle of claim 1 , wherein the machine-learned motion planning model comprises one or more first convolutional network layers that are configured to:
generate one or more intermediate representations associated with at least one of an object detection or an objection prediction based at least in part on the sensor data and the map data.
3. The autonomous vehicle of claim 2 , wherein:
the one or more intermediate representations include one or more bounding boxes associated with the object detection and one or more motion predictions associated with the object prediction.
4. The autonomous vehicle of claim 3 , wherein the machine-learned motion planning model comprises one or more second convolutional network layers that are configured to:
generate the cost volume based at least in part on the sensor data and the map data.
5. The autonomous vehicle of claim 4 , wherein:
the one or more first convolutional network layers are optimized based on an output of the one or more second convolutional network layers.
6. The autonomous vehicle of claim 5 , wherein:
the machine-learned motion planning model is trained based at least in part on multitask training with supervision for perception and motion planning.
7. The autonomous vehicle of claim 6 , wherein:
the machine-learned motion planning model is trained using a total loss function that includes a perception loss component and a motion planning loss component.
8. The autonomous vehicle of claim 7 , wherein the perception loss component includes:
a classification loss associated with distinguishing a vehicle from a background; and
a regression loss associated with generating object bounding boxes.
9. The autonomous vehicle of claim 8 , wherein:
the motion planning loss component is generated based at least in part on one or more human-driven trajectories.
10. The autonomous vehicle of claim 1 , wherein the machine-learned motion planning model is configured to generate the cost volume with a backbone network that outputs the cost volume to a trajectory generator that selects the target trajectory.
11. The autonomous vehicle of claim 2 , wherein the output of the machine-learned motion planning model is a first output, the operations further comprising:
receiving the one or more intermediate representations as a second output of the machine-learned motion planning model.
12. The autonomous vehicle of claim 2 , wherein:
the machine-learned motion planning model is jointly trained for motion planning and generating the intermediate representations based on motion planning optimization.
13. The autonomous vehicle of claim 1 , wherein:
the operations further comprise generating one or more vehicle control signals for the autonomous vehicle based at least in part on the target trajectory.
14. The autonomous vehicle of claim 1 , wherein using the waypoints and the plurality of timesteps to obtain cost data from the cost volume for determining the respective trajectory score comprises:
obtaining cost values for the waypoints by, for a respective timestep of the plurality of timesteps that is associated with a respective waypoint of the waypoints, applying a respective filter to the cost volume and obtaining, from the filtered cost volume, a respective value associated with the respective waypoint, and
combining the cost values to obtain the respective trajectory score.
15. The autonomous vehicle of claim 1 , wherein the machine-learned motion planning model is configured to select the target trajectory by optimizing at least one of the plurality of potential trajectories for the autonomous vehicle based at least in part on the cost volume.
16. One or more non-transitory computer-readable media that store a machine-learned motion planning model, the machine-learned motion planning model configured to:
receive sensor data and map data associated with an environment external to an autonomous vehicle;
generate, using the sensor data and the map data, a cost volume that indexes, at each of a plurality of timesteps, cost data associated with a plurality of locations in the environment; and
select a target trajectory for the autonomous vehicle from a plurality of potential trajectories based on a comparison of the plurality of potential trajectories using the cost data indexed in the cost volume by, for a respective potential trajectory corresponding to waypoints respectively associated with the plurality of timesteps, using the waypoints and the plurality of timesteps to obtain cost data from the cost volume for determining a respective trajectory score.
17. The one or more non-transitory computer-readable media of claim 16 , wherein:
the machine-learned motion planning model comprises:
one or more first convolutional network layers configured to generate one or more intermediate representations based at least in part on the sensor data and the map data; and
one or more second convolutional network layers configured to generate the cost volume based at least in part on the sensor data and the map data.
18. A computer-implemented method of motion planning for an autonomous vehicle, the method comprising:
obtaining sensor data and map data associated with an environment external to the autonomous vehicle;
generating, using a machine-learned motion planning model and based on the sensor data and the map data, a cost volume that indexes, at each of a plurality of timesteps, cost data associated with a plurality of locations in the environment; and
selecting, using the machine-learned motion planning model, a target trajectory for the autonomous vehicle from a plurality of potential trajectories based on a comparison of the plurality of potential trajectories using the cost data indexed in the cost volume by, for a respective potential trajectory corresponding to waypoints respectively associated with the plurality of timesteps, using the waypoints and the plurality of timesteps to obtain cost data from the cost volume for determining a respective trajectory score.
19. The computer-implemented method of claim 18 , wherein selecting the target trajectory for the autonomous vehicle comprises:
comparing a plurality of trajectory scores respectively associated with the plurality of potential trajectories.
20. The autonomous vehicle of claim 1 , wherein the cost volume comprises precomputed scores for the plurality of locations at the plurality of timesteps.