Joint training of a narrow field of view sensor with a global map for broader context

1. A method comprising:

receiving, from a sensor on a robotic device, a captured image representative of an environment of the robotic device when the robotic device is at a location in the environment;

determining, based at least on the location of the robotic device, a rendered image representative of the environment of the robotic device;

determining, by applying at least one pre-trained machine learning model to at least the captured image and the rendered image, predicted drivability of one or more portions of the environment represented by the captured image and the rendered image, wherein the pre-trained machine learning model has been trained based on a dataset comprising pairs of captured images and associated rendered images and labels associated with drivability of one or more portions of a depicted environment of the associated captured image and the associated rendered image, wherein each such pair represents a respective environment; and

controlling, based on the predicted drivability of the one or more portions of the environment, the robotic device to navigate in the environment.

2. The method of claim 1 , wherein the rendered image is a top-down image of an area at the location of the robotic device.

3. The method of claim 1 , wherein the rendered image is a perspective image of an area at the location of the robotic device.

4. The method of claim 1 , wherein the rendered image has a wider field of view than the captured image.

5. The method of claim 1 , wherein the method further comprises:

receiving, by the robotic device from a remote device, a map of the environment, wherein the rendered image is further based on the map.

6. The method of claim 1 , wherein the method further comprises:

determining, by the robotic device, a map of an area at the location of the robotic device, wherein determining the rendered image is further based on the determined map.

7. The method of claim 6 , wherein the map includes one or more labels corresponding to the area at the location of the robotic device, and wherein the at least one pre-trained machine learning model is further applied to the one or more labels.

8. The method of claim 1 , wherein the method further comprises:

determining a point in the rendered image, wherein the point is representative of where the robotic device is located, wherein the at least one pre-trained machine learning model is further applied to the determined point.

9. The method of claim 1 , wherein the captured image and the rendered image are concatenated prior to being input into the at least one pre-trained machine learning model.

10. The method of claim 1 , wherein determining the property of the one or more portions of the environment represented by the captured image comprises:

transmitting, to a remote device, the captured image and the rendered image; and

receiving, from the remote device applying the at least one pre-trained machine learning model to the captured image and the rendered image, the property of the one or more portions of the environment represented by the captured image.

11. The method of claim 1 , wherein the at least one pre-trained machine learning model comprises an attention mechanism trained to identify a region of the rendered image such that the captured image attends to the region of the rendered image for determining the property of the one or more portions of the environment represented by the captured image.

12. The method of claim 1 , wherein one of the at least one pre-trained machine learning model is a model trained on a different environment of the robotic device and further refined based on the environment of the robotic device.

13. The method of claim 1 , wherein the method further comprises:

prior to determining the rendered image, determining the location of the robotic device in the environment.

14. A robotic device comprising:

a sensor; and

a control system configured to:

receive, from a sensor on a robotic device, a captured image representative of an environment of the robotic device when the robotic device is at a location in the environment;

determine, based at least on the location of the robotic device, a rendered image representative of the environment of the robotic device;

determine, by applying at least one pre-trained machine learning model to at least the captured image and the rendered image, predicted drivability of one or more portions of the environment represented by the captured image and the rendered image, wherein the pre-trained machine learning model has been trained based on a dataset comprising pairs of captured images and associated rendered images and labels associated with drivability of one or more portions of a depicted environment of the associated captured image and the associated rendered image, wherein each such pair represents a respective environment; and

control, based on the predicted drivability of the one or more portions of the environment, the robotic device to navigate in the environment.

15. The robotic device of claim 14 , wherein the sensor is a red green blue (RGB) camera, a red green blue depth (RGB-D) camera, a depth sensor, or a LIDAR sensor.

16. The robotic device of claim 14 , wherein the control system is further configured to:

determine a map of an area at the location of the robotic device, wherein the control system is configured to determine the rendered image based on the determined map.

17. The robotic device of claim 14 , wherein the control system is further configured to:

prior to determining the rendered image, determine the location of the robotic device in the environment.

18. A non-transitory computer readable medium comprising program instructions executable by at least one processor to cause the at least one processor to perform functions comprising:

receiving, from a sensor on a robotic device, a captured image representative of an environment of the robotic device when the robotic device is at a location in the environment;

determining, based at least on the location of the robotic device, a rendered image representative of the environment of the robotic device;

determining, by applying at least one pre-trained machine learning model to at least the captured image and the rendered image, predicted drivability of one or more portions of the environment represented by the captured image and the rendered image, wherein the pre-trained machine learning model has been trained based on a dataset comprising pairs of captured images and associated rendered images and labels associated with drivability of one or more portions of a depicted environment of the associated captured image and the associated rendered image, wherein each such pair represents a respective environment; and

controlling, based on the predicted drivability of the one or more portions of the environment, the robotic device to navigate in the environment.