Methods and apparatus for autonomous robotic control

1. A method for sensory scene processing, the method comprising:

segmenting, with a coarse observer, a visual scene into a plurality of areas based on a plurality of semantic labels;

labeling, with the coarse observer, areas within the plurality of areas with respective semantic labels from the plurality of semantic labels;

identifying, with the coarse observer, at least one high-priority area in the plurality of areas based on the semantic labels used to label the areas within the plurality of areas;

biasing a fine observer, by the coarse observer, to sample the at least one high-priority area;

in response to biasing the fine observer, sampling, by the fine observer, the at least one high-priority area;

identifying, with the fine observer, at least one object in the at least one high-priority area based at least in part on the sampling;

labeling, with the fine observer, the at least one object in the at least one high-priority area; and

storing a representation of the at least one object as metadata for the visual scene.

2. The method of claim 1 , wherein segmenting the visual scene comprises:

carving, with the coarse observer, the visual scene into sub-areas; and

extracting, with the coarse observer, a first feature vector from a first sub-area of the sub-areas.

3. The method of claim 2 , wherein segmenting the visual scene further comprises:

performing a comparison, with the coarse observer, of the first feature vector with a second feature vector from a second sub-area of the sub-areas.

4. The method of claim 1 , wherein each area of the plurality of areas is a macro-area in the visual scene.

5. The method of claim 1 , wherein biasing the fine observer comprises:

focusing, with an interface communicably coupled to the coarse observer and the fine observer, attention of the fine observer to the at least one high-priority area; and

classifying, with the fine observer, the at least one object in the at least one high-priority area.

6. The method of claim 5 , further comprising:

combining a representation of each area in the plurality of areas and a representation of the at least one object in the at least one high-priority area as metadata for the visual scene.

7. The method of claim 1 , wherein the at least one object is a first object, and further comprising:

classifying, with the coarse observer, each area in the plurality of areas;

focusing, with an interface communicably coupled to the coarse observer and the fine observer, attention of the fine observer to a first area in the plurality of areas, wherein the first area is different than the at least one high-priority area; and

classifying, with the fine observer, a second object in the first area.

8. The method of claim 1 , further comprising:

applying unsupervised learning to the coarse observer; and

categorizing a first subset of areas in the plurality of areas into a plurality of categorized areas based at least in part on the unsupervised learning.

9. The method of claim 8 , further comprising:

applying supervised learning to the coarse observer to associate a second subset of areas of the plurality of areas with user-accessible semantic labels; and

training the coarse observer based at least in part on the plurality of categorized areas and the user-accessible semantic labels.

10. The method of claim 9 , further comprising:

associating, with the coarse observer during and after the training, each area in the plurality of areas with at least one of: a measure of familiarity, a measure of typicality, or a measure of task-based relevance.

11. The method of claim 1 , further comprising:

training the fine observer based on a dataset of a plurality of labeled objects.

12. A system for sensory processing, the system comprising:

a coarse observer to:

segment a visual scene into a plurality of areas based on a plurality of semantic labels,

label areas within the plurality of areas with respective semantic labels from the plurality of semantic labels,

identify at least one high-priority area in the plurality of areas based on the semantic labels used to label the areas within the plurality of areas, and

bias a fine observer to sample the at least one high-priority area;

a fine observer, communicably coupled to the coarse observer, to:

in response to biasing the fine observer, sample the at least one high-priority area to generate at least one sampled high-priority area,

identify at least one object in the at least one sampled high-priority area, and

label at least one object in the at least one sampled high priority area; and

a memory, communicably coupled to the fine observer and the coarse observer, to store a representation of the at least one object as metadata for the visual scene.

13. The system of claim 12 , wherein the coarse observer is further configured to:

carve the visual scene into sub-areas; and

extract a first feature vector from a first sub-area of the sub-areas.

14. The system of claim 13 , wherein the coarse observer is further configured to:

perform a comparison of the first feature vector with a second feature vector from a second sub-area of the sub-areas.

15. The system of claim 12 , wherein each area of the plurality of areas is a macro-area in the visual scene.

16. The system of claim 12 , further comprising:

an interface, communicably coupled to the coarse observer and the fine observer, to focus attention of the fine observer to the at least one high-priority area, and

wherein the fine observer is further configured to classify the at least one object in the at least one high priority area.

17. The system of claim 16 , wherein the coarse observer is further configured to combine the representation of each area in the plurality of areas and the representation of the at least one object in the at least one high-priority area as metadata for the visual scene.

18. The system of claim 12 , further comprising:

an interface, communicably coupled to the coarse observer and the fine observer, to focus attention of the fine observer to a first area in the plurality of areas, wherein the first area is different than the at least one high-priority area, and wherein

the coarse observer is further configured to classify each area in the plurality of areas, and

the fine observer is further configured to classify another object in the first area.

19. The system of claim 12 , wherein the coarse observer is further configured to:

implement unsupervised learning; and

categorize a first subset of areas in the plurality of areas into a plurality of categorized areas based at least in part on the unsupervised learning.

20. The system of claim 19 , wherein the coarse observer is further configured to implement supervised learning to associate a second subset of areas of the plurality of the plurality of areas with user-accessible semantic labels, and further comprising:

a processor, communicably coupled to the coarse observer, to train the coarse observer based at least in part on the plurality of categorized areas and the user-accessible semantic labels.

21. The method of claim 20 , wherein, during and after the training, the coarse observer is further configured to associate each area in the plurality of areas with at least one of: a measure of familiarity, a measure of typicality, and a measure of task-based relevance.

22. A method for sensory scene processing, the method comprising:

segmenting, with a coarse observer, a visual scene into a plurality of areas;

labeling, with the coarse observer, the plurality of areas;

identifying, with the coarse observer, at least one high-priority area and at least one low-priority area in the plurality of areas;

biasing a fine observer, by the coarse observer, to scan the at least one high-priority area;

in response to biasing the fine observer, sampling, by the fine observer, the at least one high-priority area at a first time and the at least one low-priority area at a second time after the first time;

identifying, with the fine observer, at least one object in the at least one high-priority area based at least in part on the sampling;

labeling, with the fine observer, the at least one object in the at least one high-priority area; and

storing a representation of the at least one object as metadata for the visual scene.