Perception and motion prediction for autonomous devices

1. A computer-implemented method of perception and motion forecasting, the computer-implemented method comprising:

generating a plurality of temporal instance representations respectively associated with differences in an appearance and a motion of one or more objects over past time intervals, wherein the plurality of temporal instance representations respectively comprise one or more appearance features of the one or more objects at the past time intervals and one or more motion features of one or more objects at the past time intervals;

determining, based at least in part on the plurality of temporal instance representations and current detections of a plurality of objects comprising the one or more objects, one or more past paths of the one or more objects over the past time intervals and one or more candidate paths of the plurality of objects over a plurality of time intervals comprising a current time interval and at least one of the past time intervals;

determining one or more predicted paths of the plurality of objects based at least in part on one or more machine-learned models, the one or more machine-learned models utilizing the one or more past paths and the one or more candidate paths to infer the one or more predicted paths; and

generating path data comprising information associated with the one or more predicted paths for the plurality of objects respectively.

2. The computer-implemented method of claim 1 , wherein the one or more past paths comprise at least one null path, and wherein the determining based at least in part on the plurality of temporal instance representations and the current detections of the plurality of objects comprising the one or more objects comprises:

determining, based at least in part on one or more comparisons of the plurality of objects to the one or more objects, whether the plurality of objects includes one or more newly detected objects not included in the one or more objects from the past time intervals; and

associating the one or more newly detected objects with the at least one null path.

3. The computer-implemented method of claim 1 , wherein the generating the plurality of temporal instance representations comprises:

obtaining data associated with the motion of the one or more objects over the past time intervals from an object path memory; and

obtaining data associated with the appearance of the one or more objects over the past time intervals from an appearance memory that is different from the object path memory.

4. The computer-implemented method of claim 1 , wherein the generating the plurality of temporal instance representations comprises:

generating based at least in part on a plurality of machine-learned feature extraction models and multi-sensor data, a plurality of feature maps associated with the appearance and the motion of the one or more objects over the one or more past time intervals, wherein the multi-sensor data is based at least in part on sensor outputs from a plurality of different types of sensors; and

generating the plurality of temporal instance representations based at least in part on the plurality of feature maps.

5. The computer-implemented method of claim 4 , wherein the multi-sensor data comprises one or more light detection and ranging (LiDAR) sweeps, map data comprising information associated with one or more locations in an environment comprising the one or more objects, or one or more images comprising the one or more objects.

6. The computer-implemented method of claim 1 , wherein the plurality of temporal instance representations respectively comprise a concatenation of one or more appearance features and one or more motion features respectively associated with the appearance and the motion of the one or more objects over the past time intervals.

7. The computer-implemented method of claim 1 , wherein a number of the one or more candidate paths is at least as great as a combination of a number of the one or more past paths and a number of the current detections of the plurality of objects.

8. The computer-implemented method of claim 1 , wherein the determining the one or more predicted paths of the plurality of objects based at least in part on one or more machine-learned models comprises:

determining a plurality of matching scores corresponding to the plurality of temporal instance representations, wherein the plurality of matching scores are respectively based at least in part on differences between the appearance and the motion of the plurality of objects over the one or more past paths and the appearance and the motion of the plurality of objects over the one or more candidate paths; and

determining the one or more predicted paths based at least in part on the plurality of matching scores associated with a least amount of difference in the appearance and the motion of the plurality of objects.

9. The computer-implemented method of claim 1 , wherein the one or more machine-learned models are configured to respectively compare the appearance and the motion of the plurality of objects along the one or more past paths at the past time intervals to the appearance and the motion of the plurality of objects along the one or more candidate paths at the past time intervals.

10. The computer-implemented method of claim 1 , wherein the one or more machine-learned models are trained based at least in part on minimization of a loss associated with one or more differences between one or more predicted training paths and one or more ground-truth paths, wherein the one or more predicted training paths are generated using training data and the one or more machine-learned models, and wherein the training data comprises a plurality of training temporal instance representations and a plurality of training object detections.

11. The computer-implemented method of claim 10 , wherein the loss is based at least in part on a loss function associated with a detection loss, a matching loss, a confidence score loss, a refinement loss, or a prediction loss.

12. The computer-implemented method of claim 10 , wherein the loss is inversely correlated with similarity of the one or more predicted training paths relative to the one or more ground-truth paths.

13. The computer-implemented method of claim 1 , wherein the determining the one or more predicted paths of the plurality of objects based at least in part on one or more machine-learned models comprises:

determining for the one or more candidate paths, and based at least in part on the plurality of temporal instance representations and the one or more machine-learned models comprising a machine-learned refinement model, one or more confidence scores, one or more path refinements, and one or more candidate predicted paths;

generating one or more refined candidate paths based at least in part on the one or more candidate predicted paths and the one or more path refinements;

ranking the one or more refined candidate paths based at least in part on the one or more confidence scores; and

determining the one or more predicted paths based at least in part on the ranking of the one or more refined candidate paths.

14. The computer-implemented method of claim 13 , wherein the one or more confidence scores are associated with a respective estimated accuracy of the one or more candidate predicted paths, and wherein the one or more path refinements comprise adjustments of bounding boxes associated with the appearance of the plurality of objects along the one or more candidate paths.

15. A computing system comprising:

one or more processors;

a memory comprising one or more tangible non-transitory computer-readable media, the memory storing computer-readable instructions for execution by the one or more processors that cause the computing system to perform operations comprising:

generating a plurality of temporal instance representations respectively associated with differences in an appearance and a motion of one or more objects over past time intervals, wherein the plurality of temporal instance representations respectively comprise one or more appearance features of the one or more objects at the past time intervals and one or more motion features of one or more objects at the past time intervals;

determining, based at least in part on the plurality of temporal instance representations and current detections of a plurality of objects comprising the one or more objects, one or more past paths of the one or more objects over the past time intervals and one or more candidate paths of the plurality of objects over a plurality of time intervals comprising a current time interval and at least one of the past time intervals;

determining one or more predicted paths of the plurality of objects based at least in part on one or more machine-learned models, the one or more machine-learned models utilizing the one or more past paths and the one or more candidate paths to infer the one or more predicted paths; and

generating path data comprising information associated with the one or more predicted paths for the plurality of objects respectively.

16. The computing system of claim 15 , wherein determining the one or more predicted paths of the plurality of objects based at least in part on one or more machine-learned models comprises determining a respective confidence score associated the one or more predicted paths.

17. The computing system of claim 15 , wherein the one or more appearance features comprise colors, intensities, textures, or edges of each of the one or more objects, and wherein the one or more motion features comprise one or more locations of each of the one or more objects or one or more headings of each of the one or more objects.

18. An autonomous vehicle comprising:

one or more processors;

a memory comprising one or more tangible non-transitory computer-readable media, the memory storing computer-readable instructions for execution by the one or more processors that cause the computing system to perform operations comprising:

generating a plurality of temporal instance representations respectively associated with differences in an appearance and a motion of one or more objects over past time intervals, wherein the plurality of temporal instance representations respectively comprise one or more appearance features of the one or more objects at the past time intervals and one or more motion features of one or more objects at the past time intervals;

determining, based at least in part on the plurality of temporal instance representations and current detections of a plurality of objects comprising the one or more objects, one or more past paths of the one or more objects over the past time intervals and one or more candidate paths of the plurality of objects over a plurality of time intervals comprising a current time interval and at least one of the past time intervals;

determining one or more predicted paths of the plurality of objects based at least in part on one or more machine-learned models, the one or more machine-learned models utilizing the one or more past paths and the one or more candidate paths to infer the one or more predicted paths; and

generating path data comprising information associated with the one or more predicted paths for the plurality of objects respectively.

19. The autonomous vehicle of claim 18 , wherein the path data is part of an input to a motion planning system of the autonomous vehicle.

20. The autonomous vehicle of claim 18 , further comprising:

controlling one or more vehicle systems of the autonomous vehicle based at least in part on the path data.