Self-aware visual-textual co-grounded navigation agent
An agent for navigating a mobile automated system is disclosed herein. The navigation agent receives a navigation instruction and visual information for one or more observed images. The navigation agent is provided or equipped with self-awareness, which provides or supports the following abilities: identifying which direction to go or proceed by determining the part of the instruction that corresponds to the observed images (visual grounding), and identifying which part of the instruction has been completed or ongoing and which part is potentially needed for the next action selection (textual grounding). In some embodiments, the navigation agent applies regularization to ensures that the grounded instruction can correctly be used to estimate the progress made towards the navigation goal (progress monitoring).
1. A computing device of a mobile automated system comprising:
a memory containing machine readable medium storing machine executable code; and
one or more processors coupled to the memory and configurable to execute the machine executable code to cause the one or more processors to:
receive training data comprising one or more training navigation instruction samples;
generate, by a first attention module, an instruction grounding from the training navigation instruction sample for instructing the mobile automated system to navigate an environment in which the mobile automated system is located;
generate, by a second attention module, a visual grounding from one or more training images observed for the environment as the mobile automated system is navigating therethrough at the first time instance;
generate, by a neural model, a progress monitor probability in response to an input of the instruction grounding and the visual grounding;
generate an action probability of navigation directions based on a hidden state of the neural model and the instruction grounding;
train the neural model based at least in part on a first cross entropy between the progress monitor probability and a target progress label, and a second cross-entropy between the action probability and a ground-truth navigable direction; and
generate a navigational action for the mobile automated system to navigate the environment by the trained neural model.
2. The computing device of claim 1 , wherein the machine executable code further causes the one or more processors to generate the instruction grounding based at least in part on the training navigation instruction sample, the instruction grounding identifying which part of the training navigation instruction sample has been completed by the mobile automated system and which part of the training navigation instruction sample is outstanding.
3. The computing device of claim 1 , wherein the machine executable code further causes the one or more processors to generate the visual grounding based at least in part on the visual information contained in the training images, the visual grounding identifying a direction in which the mobile automated system should proceed.
4. The computing device of claim 1 , wherein the ground-truth progress label is computed as a normalized distance from a current viewpoint to a goal.
5. The computing device of claim 1 , wherein the loss objective is computed as a linear combination of the first cross-entropy and the second cross-entropy with a weighting parameter adjusting a ratio between the first cross-entropy and the second cross-entropy.
6. The computing device of claim 1 , wherein the machine executable code further causes the one or more processors to monitor progress of navigation of the automated system to ensure that the instruction grounding accurately reflects the navigation progress.
7. The computing device of claim 1 , wherein the machine executable code causes the one or more processors to:
generate an encoder context based on the instruction grounding and the visual grounding; and
generate the action probability for the mobile automated system using the encoder context.
8. The computing device of claim 1 , wherein the machine executable code causes the one or more processors to perform a natural language processing task on the training navigation instruction sample.
9. The computing device of claim 1 , wherein the machine executable code causes the one or more processors to identify a navigable direction with the highest correlation to the instruction grounding.
10. The computing device of claim 1 , wherein the machine executable code causes the one or more processors to:
identify a plurality of directions in which the mobile automated system can navigate; and
for each identified navigable direction, generate a respective probability.
11. A method of training a neural model at a computing device of a mobile automated system for monitoring navigation progress, the method comprising:
receiving, at the computing device, training data comprising one or more training navigation instruction samples;
generating, by a first attention module, an instruction grounding from the training navigation instruction sample for instructing the mobile automated system to navigate an environment in which the mobile automated system is located;
generating, by a second attention module, a visual grounding from one or more training images observed for the environment as the mobile automated system is navigating therethrough at the first time instance;
generating, by a neural model, a progress monitor probability in response to an input of the instruction grounding and the visual grounding;
generating an action probability of navigation directions based on a hidden state of the neural model and the instruction grounding;
training the neural model based at least in part on a first cross entropy between the progress monitor probability and a target progress label, and a second cross-entropy between the action probability and a ground-truth navigable direction; and
generating a navigational action for the mobile automated system to navigate the environment by the trained neural model.
12. The method of claim 11 , further comprising:
generating the instruction grounding based at least in part on the training navigation instruction sample, the instruction grounding identifying which part of the training navigation instruction sample has been completed by the mobile automated system and which part of the training navigation instruction sample is outstanding.
13. The method of claim 11 , further comprising:
generating the visual grounding based at least in part on the visual information contained in the training images, the visual grounding identifying a direction in which the mobile automated system should proceed.
14. The method of claim 11 , wherein the ground-truth progress label is computed as a normalized distance from a current viewpoint to a goal.
15. The method of claim 11 , wherein the loss objective is computed as a linear combination of the first cross-entropy and the second cross-entropy with a weighting parameter adjusting a ratio between the first cross-entropy and the second cross-entropy.
16. The method of claim 11 , comprising monitoring progress of navigation of the automated system to ensure that the instruction grounding accurately reflects the navigation progress.
17. The method of claim 11 , wherein the action probability is computed by:
generating an encoder context based on the instruction grounding and the visual grounding; and
generating the action probability for the mobile automated system using the encoder context.
18. The method of claim 11 , wherein computing the action probability comprises identifying a navigable direction with the highest correlation to the instruction grounding.
19. The method of claim 11 , wherein computing the action probability comprises:
identifying a plurality of directions in which the mobile automated system can navigate; and
for each identified navigable direction, generating a respective probability.
20. A non-transitory machine-readable medium comprising executable code which when executed by one or more processors associated with a computing device of a mobile automated system are adapted to cause the one or more processors to perform a method comprising:
receiving, at the computing device, training data comprising one or more training navigation instruction samples;
generating, by a first attention module, an instruction grounding from the training navigation instruction sample for instructing the mobile automated system to navigate an environment in which the mobile automated system is located;
generating, by a second attention module, a visual grounding from one or more training images observed for the environment as the mobile automated system is navigating therethrough at the first time instance;
generating, by a neural model, a progress monitor probability in response to an input of the instruction grounding and the visual grounding;
generating an action probability of navigation directions based on a hidden state of the neural model and the instruction grounding;
train the neural model based at least in part on a first cross entropy between the progress monitor probability and a target progress label, and a second cross-entropy between the action probability and a ground-truth navigable direction; and
generating a navigational action for the mobile automated system to navigate the environment by the trained neural model.