Systems and methods for improving generalization in visual navigation

1. A training system for training a trained model for use by a navigating robot to perform visual navigation, the training system comprising:

memory including N base virtual training environments, each of the N base virtual training environments including a field of view at a location within an indoor space,

wherein N is an integer greater than 1;

a randomization module configured to generate N varied virtual training environments based on the N base virtual training environments, respectively, by varying at least one characteristic of the respective N base virtual training environments; and

a training module configured to train the trained model for use by the navigating robot to perform visual navigation based on a training set including:

the N base virtual training environments; and

the N varied virtual training environments,

wherein the training includes the training module:

first training the trained model using reinforcement learning to optimize a proximal policy optimization (PPO) objective based on a reward determined based on whether or not a goal location is reached; and

second training the trained model using regularization based on minimizing a penalty term that acts as a constraint on the PPO objective.

2. The training system of claim 1 wherein the at least one characteristic includes a texture of a wall.

3. The training system of claim 1 wherein the at least one characteristic includes a color of a wall.

4. The training system of claim 1 wherein the at least one characteristic includes a lighting characteristic.

5. The training system of claim 1 wherein the at least one characteristic includes a location of an object.

6. The training system of claim 1 wherein the training module is configured to train the trained model using a convolutional neural network (CNN).

7. The training system of claim 1 wherein the training module is configured to train the trained model using invariance regularization (IR).

8. The training system of claim 1 wherein the training includes iteratively switching between the first and second training.

9. The training system of claim 1 wherein the trained model is configured to, based on an input image from a camera, generate an output indicative of an instruction to one of:

move the navigating robot forward by a predetermined distance;

turn the navigating robot right by a first predetermined angle;

turn the navigating robot left by a second predetermined angle; and

not move the navigating robot.

10. The training system of claim 9 wherein the trained model is configured to generate the output based on a red green blue (RGB) input image from the camera.

11. The training system of claim 9 wherein the trained model is configured to generate the output based on a red green blue-depth (RGB-D) input image from the camera.

12. The training system of claim 9 wherein the trained model is configured to generate the output based on a grayscale input image from the camera.

13. The training system of claim 9 wherein the trained model is configured to generate the output based on a grayscale-depth (grayscale-D) input image from the camera.

14. The training system of claim 1 , wherein the indoor space is within a building.

15. A method for training a trained model for use by a navigating robot to perform visual navigation, the method comprising:

by one or more processors, generating N varied virtual training environments based on N base virtual training environments, respectively, by varying at least one characteristic of the respective N base virtual training environments,

wherein each of the N base virtual training environments and each of the N varied virtual training environments includes a field of view at a location within an indoor space, and

wherein N is an integer greater than 1; and

by the one or more processors, training the trained model for use by the navigating robot to perform visual navigation based on a training set including:

the N base virtual training environments; and

the N varied virtual training environments,

wherein the training includes:

first training the trained model using reinforcement learning to optimize a proximal policy optimization (PPO) objective based on a reward set determined on whether or not a goal location is reached; and

second training the trained model using regularization based on minimizing a penalty term that acts as a constraint on the PPO objective.

16. The method of claim 15 wherein the at least one characteristic includes one of:

a texture of a wall;

a color of a wall;

a lighting characteristic; and

a location of an object.

17. The method of claim 15 wherein the training includes training the trained model using a convolutional neural network (CNN).

18. The method of claim 15 wherein the training includes training the trained model using invariance regularization (IR).

19. The method of claim 15 wherein the trained model is configured to, based on an input image from a camera, generate an output indicative of an instruction to one of:

move the navigating robot forward by a predetermined distance;

turn the navigating robot right by a first predetermined angle;

turn the navigating robot left by a second predetermined angle; and

not move the navigating robot.

20. The method of claim 19 wherein the trained model is configured to generate the output based on one of:

a red green blue (RGB) input image from the camera;

a red green blue-depth (RGB-D) input image from the camera;

a grayscale input image from the camera; and

a grayscale-depth (grayscale-D) input image from the camera.

21. A navigating robot, comprising:

a camera configured to capture images within a field of view in front of the navigating robot,

wherein the field of view is less than 360 degrees around the navigating robot;

a model that is:

trained based on:

N base virtual training environments at locations within an indoor space; and

N varied virtual training environments generated by varying at least one characteristic of the N base virtual training environments, respectively,

wherein N is an integer greater than 1, and

the training including:

first training the model using reinforcement learning to optimize a proximal policy optimization (PPO) objective based on a reward determined based on whether or not a goal location is reached; and

second training the model using regularization based on minimizing a penalty term that acts as a constraint on the PPO objective; and

configured to generate an output based on one of the images from the camera;

one or more propulsion devices configured to propel the navigating robot; and

a control module that actuates one or more of the one or more propulsion devices based on the output of the model.

22. The navigating robot of claim 21 wherein the model is configured to set the output to, based on the one of the images from the camera, one of a group consisting of a first state, a second state, a third state, and a fourth state; and

wherein the control module is configured to:

when the output is in the first state, actuate one or more of the one or more propulsion devices to move the navigating robot forward by a predetermined distance;

when the output is in the second state, actuate one or more of the one or more propulsion devices to turn the navigating robot right by a first predetermined angle;

when the output is in the third state, actuate one or more of the one or more propulsion devices to turn the navigating robot left by a second predetermined angle; and

when the output is in the fourth state, not actuate any of the one or more propulsion devices to not move the navigating robot.

23. The navigating robot of claim 21 wherein the navigating robot is an autonomous vehicle.